A method and system for pulmonary embolism detection and segmentation based on inter-hospital distance compensation

Abstract

This invention provides a pulmonary embolism detection and segmentation method based on intercalation distance compensation, belonging to the field of computer medical image processing technology. The method includes: pre-labeling and preprocessing a lung CT image set to obtain pre-labeled labels and a standard image set; obtaining thrombus labels from the pre-labeled labels; integrating the thrombus labels and the standard image set and inputting them into a 3D target detection network model for training to obtain a detection training model; performing detection and recognition through multiple detection training models to obtain multiple preliminary detection results and fusing them into a recognition result; integrating the thrombus labels, the standard image set, and the lung CT image set and inputting them into a semi-supervised semantic segmentation model for training to obtain a final segmentation model; and using the final segmentation model and an intercalation distance compensation mechanism to segment and correct the recognition result to obtain a merged result. This method intelligently identifies pulmonary embolism through deep learning and image processing, improving the detection accuracy of pulmonary embolism.

Description

Technical Field

[0001] This invention relates to the field of computer medical image processing technology, and in particular to a method and system for detecting and segmenting pulmonary embolism based on intercalation distance compensation. Background Technology

[0002] Pulmonary embolism is a common and life-threatening condition, usually caused by a blood clot blocking the pulmonary artery or its branches, leading to impaired blood supply to the lungs. The clinical presentation of acute pulmonary embolism is diverse, ranging from asymptomatic to sudden dyspnea, chest pain, hypotension, and even sudden death. Early identification of pulmonary embolism is crucial, but traditional diagnostic methods, such as lung ventilation / perfusion scanning, CT pulmonary angiography (CTPA), and echocardiography, rely on interpretation by experienced radiologists, limiting the speed and accuracy of diagnosis.

[0003] In recent years, intelligent analysis methods have been widely researched and applied in the field of medical imaging. Automated extraction of vascular centerlines, as one of the key technologies, has become an important means to improve the accuracy and efficiency of pulmonary embolism detection. By segmenting pulmonary vessels, the structure of the pulmonary artery and its branches can be accurately depicted, providing rich spatial information for pulmonary embolism detection. Furthermore, the extraction of vascular centerlines can provide the geometric structural features of pulmonary vessels, helping to segment and locate thrombus areas.

[0004] Currently, the diagnosis of pulmonary embolism has the following shortcomings: 1. Complex image interpretation process: Although existing imaging examinations such as CTPA can show the location and size of the embolism well, the interpretation of the images requires experienced doctors, and misjudgment may occur due to visual fatigue or subjective differences during the interpretation process; 2. Lack of intelligent analysis tools: Although imaging examinations can provide a lot of information about the pulmonary artery and surrounding tissues, there is a lack of widely used intelligent tools to automatically analyze and process these images, making it difficult to obtain accurate diagnostic results in a short time.

[0005] Similarly, existing AI-based assisted diagnostic methods also have the following shortcomings: 1. High cost of medical data annotation: Medical data annotation, especially in diseases like pulmonary embolism, often requires precise pixel-level annotation by doctors or professionals. Due to the massive amount of medical image data and the high annotation requirements, the annotation cost and time consumption are enormous, limiting data availability. Furthermore, due to the diversity and complexity of diseases, the annotated data is often insufficient to cover all lesion conditions, leading to unstable model performance on different lesion samples. 2. Insufficient accuracy of existing segmentation models: Many current medical image segmentation models, such as those based on convolutional neural networks (CNNs), while achieving good results in some tasks, still face many challenges in pulmonary embolism detection and segmentation. For example, the morphology and size of thrombi vary greatly, and models may struggle to adapt to these variations, resulting in inaccurate segmentation results. In addition, existing segmentation models often perform poorly when handling details, complex boundaries, and small thrombus regions, making accurate segmentation difficult. Therefore, designing a pulmonary embolism detection and segmentation method and system based on intercalary distance compensation is essential. Summary of the Invention

[0006] The purpose of this invention is to provide a pulmonary embolism detection and segmentation method and system based on intercalary distance compensation, so as to intelligently identify the pulmonary embolism region through deep learning and image processing technology, and improve the detection accuracy of pulmonary embolism by utilizing the vascular centerline.

[0007] To achieve the above objectives, the present invention provides the following solution:

[0008] A pulmonary embolism detection and segmentation method based on intercalation distance compensation includes the following steps:

[0009] The lung CT image set was pre-annotated to obtain pre-annotated labels; the pre-annotation included: manual annotation and pixel-level annotation;

[0010] Thrombosis labels are obtained through pre-labeled tags;

[0011] Preprocessing operations were performed on the lung CT image set to obtain a standard image set. The preprocessing operations included: adjusting window width and window level, adjusting image spacing, adjusting image orientation cosine matrix, image denoising, and hybrid cropping.

[0012] The thrombus label and standard image set are integrated into a detection dataset. The detection dataset is then input into a 3D target detection network model for model training, resulting in multiple detection training models.

[0013] Multiple detection training models are used to detect and identify the detection dataset, resulting in multiple preliminary detection results. These preliminary detection results are then fused together to obtain the final identification result.

[0014] Thrombosis labels, standard image sets, and lung CT image sets are integrated into a segmentation dataset. The semi-supervised semantic segmentation model is trained using the segmentation dataset to obtain the final segmentation model.

[0015] The recognition results are segmented using the final segmentation model to obtain the segmentation result;

[0016] The segmentation result is corrected by the intermediary distance compensation mechanism to obtain the merged result.

[0017] Optionally, the thrombus label is obtained through pre-labeled labels. Specifically, the coordinates of the lung CT image are detected according to the pre-labeled labels to obtain the thrombus bounding box, and the thrombus label is obtained according to the coordinate scalars of the six corners of the thrombus bounding box; the thrombus bounding box has a cuboid structure.

[0018] Optionally, thrombus labels and a standard image set are integrated into a detection dataset. This detection dataset is then input into a 3D object detection network model for training, resulting in multiple detection training models, including:

[0019] The detection dataset is divided into a detection training set and a detection validation set.

[0020] The 3D object detection network model was trained using the detection training set to obtain several preliminary training models.

[0021] Thrombosis features are obtained by extracting features from the detection training set using multi-scale feature fusion technology.

[0022] Anchor frames of various sizes are set according to the characteristics of thrombi, and the thrombus area is determined by the anchor frames;

[0023] The initial training model is optimized by using the thrombus region and the first loss function to obtain the detection training model.

[0024] Optionally, the first loss function includes: location loss and confidence loss; the formula for calculating the first loss function is: Where GIoU is the first loss function, IoU is the IoU loss function, A is the optimal anchor box, B is the thrombus label, C is the minimum closed box formed by A and B, |C| is the area of ​​the minimum closed box, and |C\(A∪B)| is the area of ​​C after removing the union of A and B.

[0025] Optionally, the thrombosis labels, standard image set, and lung CT image set are integrated into a segmentation dataset. The semi-supervised semantic segmentation model is trained using the segmentation dataset to obtain the final segmentation model. Specifically, the segmentation dataset is input into the semi-supervised semantic segmentation model, and the model is trained using consistency regularization loss, enhanced perturbation strategy, pseudo-label generation strategy, cross-validation, and a second loss function to obtain the final segmentation model.

[0026] Optionally, the expression for the second loss function is: L = L l +λL u Where L is the value of the second loss function, L l For the labeling loss function, L u Here, λ represents the unlabeled loss function, and λ is the weight coefficient.

[0027] The expression for the label loss function is: Among them, B l Let H be the number of labeled images, and H be the hard cross-entropy loss function. Let be the predicted value for the i-th image. The true label for the i-th image;

[0028] The expression for the unlabeled loss function is: Among them, B u Number of unlabeled images For the i-th pseudo-label of the weakly enhanced image, Let τ be the i-th pseudo-label of the strongly enhanced image, and τ be a preset confidence threshold. This is a hard, unique hot tag.

[0029] Optionally, the identification results are segmented using the final segmentation model to obtain the segmentation results. Specifically, candidate regions containing pulmonary embolism in the identification results are selected as the input of the final segmentation model. High-confidence pseudo-labels are generated based on the unlabeled parts in the input of the final segmentation model, and the identification results are segmented using the high-confidence pseudo-labels to obtain the segmentation results.

[0030] Optionally, the segmentation result is corrected through an intermediary distance compensation mechanism to obtain the merging result, including:

[0031] The inter-segment distance between the segmentation results is calculated using a search algorithm;

[0032] The two thrombi in the segmentation results are judged based on the comparison between the intercalation distance and the preset threshold, and the judgment result is: the same thrombus or different thrombi.

[0033] Two thrombi that are determined to be the same thrombus are merged to obtain the merged result.

[0034] Optionally, the intercalary distance between the segmentation results is calculated using a search algorithm. Specifically, the thrombus boundary is extracted from the segmentation results, and the intercalary distance is obtained by calculating the pixel distance between the two closest points of the thrombus using the vessel centerline and the thrombus boundary.

[0035] A pulmonary embolism detection and segmentation system based on intercalation distance compensation, comprising:

[0036] The image acquisition module is used to pre-annotate the lung CT image set to obtain pre-annotated labels;

[0037] The preprocessing module is used to perform preprocessing operations on the lung CT image set to obtain a standard image set;

[0038] The detection module integrates thrombus labels and standard image sets into a detection dataset. The detection dataset is then input into a 3D target detection network model for training, resulting in multiple detection training models. These multiple training models are then used to detect and identify the dataset, yielding multiple preliminary detection results. Finally, these preliminary detection results are fused together to obtain the final recognition result.

[0039] The segmentation module integrates thrombus labels, standard image sets, and lung CT image sets into a segmentation dataset. It then trains a semi-supervised semantic segmentation model using this dataset to obtain the final segmentation model. Finally, it segments the recognition results using the final segmentation model to obtain the segmentation result.

[0040] The accuracy compensation module is used to correct the segmentation results through the intercalation distance compensation mechanism to obtain the merged result;

[0041] The visualization and report generation module is used to visualize the merged results in three dimensions and generate screening reports.

[0042] This invention discloses the following technical effects: The method and system for pulmonary embolism detection and segmentation based on intercalation distance compensation provided by this invention include: pre-annotating a lung CT image set to obtain pre-annotated labels; obtaining thrombus labels through the pre-annotated labels; performing preprocessing operations on the lung CT image set to obtain a standard image set; integrating the thrombus labels and the standard image set into a detection dataset; inputting the detection dataset into a 3D target detection network model for model training to obtain multiple detection training models; performing detection and recognition on the detection dataset through multiple detection training models to obtain multiple preliminary detection results, and fusing the multiple preliminary detection results to obtain a recognition result; integrating the thrombus labels, the standard image set, and the lung CT image set into a segmentation dataset; training a semi-supervised semantic segmentation model through the segmentation dataset to obtain a final segmentation model; segmenting the recognition result using the final segmentation model to obtain a segmentation result; and correcting the segmentation result through an intercalation distance compensation mechanism to obtain a merged result. This method intelligently identifies pulmonary embolism through deep learning and image processing, improving the detection accuracy of pulmonary embolism. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart of the pulmonary embolism detection and segmentation method of the present invention;

[0045] Figure 2 This is a flowchart of the training process for the 3D target detection network model of the present invention;

[0046] Figure 3 This is a diagram of the detection training model of the present invention;

[0047] Figure 4 This is a flowchart of the training process for the semi-supervised semantic segmentation model of the present invention;

[0048] Figure 5 This is a flowchart illustrating the segmentation result correction process of the present invention.

[0049] Figure 6 This is a first slice image for thrombus identification according to an embodiment of the present invention;

[0050] Figure 7 This is a second slice image for thrombus identification according to an embodiment of the present invention;

[0051] Figure 8 This is a third slice image for thrombus identification according to an embodiment of the present invention;

[0052] Figure 9 This is a first slice view of the missing portion in an embodiment of the present invention;

[0053] Figure 10 This is a second slice view of the missing portion in an embodiment of the present invention;

[0054] Figure 11 This is a three-dimensional view of the missing portion in an embodiment of the present invention;

[0055] Figure 12 This is a first slice of the merging result according to an embodiment of the present invention;

[0056] Figure 13 This is a second slice of the merging result according to an embodiment of the present invention;

[0057] Figure 14 This is the third slice of the merging result in an embodiment of the present invention;

[0058] Figure 15 This is a three-dimensional diagram of the merging result in an embodiment of the present invention. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0061] like Figure 1 As shown, the present invention provides a pulmonary embolism detection and segmentation method based on intercalation distance compensation, comprising the following steps:

[0062] Step 100: Pre-annotate the lung CT image set to obtain pre-annotated labels; the pre-annotation includes: simple manual annotation and pixel-level annotation.

[0063] Specifically, doctors manually annotate the first, largest, and last slices of lung CT images to identify the location of the thrombus, obtaining pre-labeled sections. The actual number of annotated slices depends on the thrombus's condition in different CT images. A small number of thrombi are annotated completely, with each slice annotated pixel-level.

[0064] It should be noted that after obtaining the patient's lung CT images, doctors only need to perform simple manual annotation on the vast majority of them. Only the remaining small portion of the CT images requires pixel-level annotation for pulmonary embolism. Compared to fully supervised semantic segmentation, where all data requires pixel-level annotation, this method significantly reduces the burden of dataset construction in semantic segmentation tasks.

[0065] Step 200: Obtain thrombus tags through pre-labeled tags.

[0066] Specifically, coordinate detection is performed on the lung CT image based on pre-labeled data to obtain a cuboid AABB thrombus bounding box. Then, the coordinates of the two diagonals of the thrombus bounding box, i.e., six scalars, are determined. Based on the obtained six scalars, a thrombus bounding box label (i.e., thrombus label) with the same data format as the lung CT image is constructed.

[0067] It should be noted that the thrombus surrounding box label is composed of the center point of a cuboid structure and its length, width and height. The process of obtaining the cuboid structure is called coordinate detection.

[0068] Step 300: Perform preprocessing operations on the lung CT image set to obtain a standard image set. Preprocessing operations include: adjusting window width and level, adjusting image spacing, adjusting image orientation cosine matrix, image denoising, and hybrid cropping.

[0069] Specifically, adjusting the window width and level involves adjusting the window width and level of the lung CT images to a uniform range for observing thrombus structures in the lung CT images. In this embodiment, the uniform window width and level are 350 / 40. Adjusting the image spacing involves standardizing the pixel spacing between different lung CT images to a uniform standard spacing, ensuring that the images input into the model have the same pixel spacing. Adjusting the image direction cosine matrix involves unifying the image direction cosine matrix to an identity matrix. Image denoising involves processing the CT images using denoising techniques such as Gaussian filtering and median filtering to reduce noise interference and improve image quality. Hybrid cropping involves randomly cutting and pasting image regions between two images while simultaneously mixing their labels to enhance the robustness and generalization ability of the model.

[0070] Step 400: Integrate the thrombus labels and standard image set into a detection dataset. Input the detection dataset into a 3D object detection network model for model training, resulting in multiple detection training models. Specifically, as follows... Figure 2 As shown, it includes:

[0071] Step 401: Divide the detection dataset into a detection training set and a detection validation set;

[0072] Specifically, cross-validation is used to reasonably divide the detection dataset into a detection training set and a detection validation set.

[0073] Step 402: Train the 3D object detection network model using the detection training set to obtain multiple preliminary training models;

[0074] Specifically, the detection dataset is input into a 3D target detection model for training to establish a mapping relationship between CT image features and thrombus locations, providing a foundation for thrombus detection. Multiple preliminary training models are generated through cross-validation, and each preliminary training model is evaluated and its performance is verified on different detection datasets.

[0075] It should be noted that the model performance is continuously monitored and evaluated during the model training process to ensure that the model has good generalization ability, effectively reducing the risk of overfitting, and enabling the model to maintain a high detection accuracy on unseen test data.

[0076] More specifically, the 3D object detection network model in this embodiment adopts the nnDetection framework. Based on a modified RetinaNet structure, the nnDetection framework can automatically perform object detection to improve the accuracy and efficiency of pulmonary embolism detection. This framework features a highly modular design, supporting automated network configuration and optimization to adapt to different medical image data. During training, the nnDetection framework can automatically configure the hyperparameters and architecture of the neural network, including adjusting parameters such as the number of layers, convolutional kernel size, and learning rate, thereby finding the optimal model structure. This automated configuration allows the model to better adapt to CT image data from different patients, improving the robustness of pulmonary embolism detection. The RetinaUNet structure serves as the base network, combining segmentation and object detection capabilities. By performing feature extraction and feature fusion on lung CT images, this nnDetection framework structure can accurately locate the detection box position of pulmonary embolism and calculate the confidence score.

[0077] Step 403: Extract features from the detection training set using multi-scale feature fusion technology to obtain thrombus features;

[0078] Specifically, the multi-scale feature fusion technique utilizes multiple convolutional and pooling layers to extract feature maps from different scales, enabling the model to identify thrombus features from details to the whole, ensuring accurate detection of thrombi of various sizes and shapes, and generating thrombus features (thrombus location and size). Lower-level feature maps retain higher spatial resolution, which helps identify small thrombi, while higher-level feature maps provide stronger semantic information, aiding in the localization of large thrombi.

[0079] Step 404: Set anchor frames of multiple sizes according to the characteristics of the thrombus, and determine the thrombus area through the anchor frames;

[0080] Specifically, based on the extracted thrombus features, anchor boxes of different sizes are set for each image region. Each anchor box can cover thrombus regions of different scales, and the three-dimensional spatial location of the thrombus is accurately located by calculating the overlap between the anchor box and the thrombus target. The position, size, and shape of the anchor boxes are adjusted based on thrombus instances in the training data. Redundant anchor boxes are removed using the non-maximum suppression (NMS) algorithm, retaining the optimal anchor boxes to ensure more accurate detection results.

[0081] It should be noted that K-means clustering algorithm is used to generate anchor boxes that are more suitable for detection, thereby improving the accuracy of anchor boxes and localization precision. By employing a multi-size anchor box design to address the diversity of thrombi (including variations in shape, size, and location), the model's detection capability is enhanced when faced with different thrombus morphologies, thus improving the accuracy of thrombus identification and localization.

[0082] Step 405: Optimize the initial training model using the thrombus region and the first loss function to obtain the detection training model.

[0083] Specifically, appropriate loss functions are introduced during training, including location loss and confidence loss. Location loss optimizes the accuracy of thrombus location, while confidence loss optimizes the reliability of thrombus detection. During training, the model iteratively optimizes to reduce these losses and improve the final detection performance.

[0084] Specifically, the confidence loss function uses the BCE loss function. The formula for calculating the first loss function is:

[0085]

[0086] Where GIoU is the loss function, IoU is the IoU loss function, A is the optimal anchor box, B is the thrombus label, C is the minimum closed box formed by A and B, |C| is the area of ​​the minimum closed box, and |C\(A∪B)| is the area of ​​the union of A and B removed from C.

[0087] It should be noted that by introducing loss function optimization techniques during the training process, and combining confidence loss and location loss, the accuracy of the model in locating thrombus regions and the precision in identifying thrombi were improved.

[0088] Furthermore, a spatial attention mechanism was introduced during the optimized training process to enhance the model's focus on thrombus regions. This mechanism enables the model to pay closer attention to key areas during detection, improving the accuracy of thrombus localization. It adaptively adjusts the model's attention distribution based on the spatial characteristics of the thrombus, thus enabling more accurate thrombus detection in complex backgrounds.

[0089] like Figure 3 As shown, the input image sequentially passes through the attention mechanism and feature extraction network of the detection training model, yielding two outputs: one output is obtained by regressing the coordinates of the anchor boxes, and the other output is a classification of the size of the thrombus or anchor box. Both outputs are supervised using the coordinates and size of the anchor boxes and loss is calculated accordingly. The feature extraction network has a built-in feature pyramid structure, which facilitates feature transfer between different feature extraction networks through convolution + normalization + activation functions, and also through skip connections.

[0090] Step 500: Use multiple detection training models to detect and identify the detection dataset to obtain multiple preliminary detection results, and then fuse the multiple preliminary detection results to obtain the identification result.

[0091] Specifically, by fusing the outputs of multiple detection training models through weighted averaging or voting mechanisms, the accuracy and robustness of detection are further improved, the impact of possible biases in the output of a single model on the results is reduced, and the ability to identify complex thrombus morphologies is enhanced.

[0092] Step 600: Integrate the thrombus labels, standard image set, and lung CT image set into a segmentation dataset. Train the semi-supervised semantic segmentation model using this dataset to obtain the final segmentation model. Specifically, as follows... Figure 4 As shown, it includes:

[0093] Step 601: Integrate the thrombus labels, standard image set, and lung CT image set into a segmentation dataset;

[0094] Specifically, a segmentation dataset containing both real and pseudo-labels is created by combining thrombus labels, a standard image set, and a lung CT image set. An adaptive pseudo-label generation strategy is used to initialize predictions on unlabeled lung CT images, thereby generating pseudo-labels for these images. The quality of the pseudo-labels is dynamically adjusted based on the model's confidence level; low-confidence pseudo-labels are assigned lower weights to prevent negative impacts on training. The anchor boxes are expanded by a certain percentage of pixels based on their size to better capture the contour information of the thrombus; simultaneously, keeping the thrombus segmentation within a smaller range allows for better segmentation results with fewer network parameters.

[0095] Step 602: Optimize the semi-supervised semantic segmentation model by performing semi-supervised learning on the segmented dataset to obtain the optimized segmentation model;

[0096] Specifically, the semi-supervised semantic segmentation model in this embodiment adopts the Unimatch framework based on the U-Net network structure. The Unimatch framework improves the accuracy and robustness of semantic segmentation by combining labeled and unlabeled data in a semi-supervised learning approach. In the diagnosis of pulmonary embolism in lung CT images, the Unimatch framework utilizes bounding box regions from target detection results for segmentation, effectively reducing segmentation computation and improving overall segmentation efficiency. The Unimatch framework employs consistency regularization and feature alignment techniques for feature learning, ensuring effective utilization of unlabeled data, and enhances the model's accurate segmentation ability of thrombus regions through multi-scale feature extraction. The Unimatch framework also uses a pseudo-label generation strategy, using high-confidence segmentation results from unlabeled data as pseudo-labels, expanding the scope of labeled data utilization, improving the model's generalization ability, and enabling it to better adapt to the pulmonary embolism morphological characteristics of different patients. The Unimatch framework inherits U-Net's powerful feature extraction and contextual information aggregation capabilities, performs refined segmentation of thrombus regions in the input CT images through an encoder-decoder structure, and further improves model performance by combining semi-supervised learning techniques. The encoder consists of multiple convolutional and pooling layers to extract multi-scale feature information from the image; the decoder restores the spatial resolution of the image through upsampling and convolutional layers, and finally outputs a segmentation result with the same size as the input image.

[0097] Step 603: Perform cross-validation on the optimized segmentation model;

[0098] Specifically, cross-validation was used to evaluate the optimized segmentation model, with the validation set consisting of CT images from different data sources to ensure the model's generalization ability. Evaluation metrics used included the Dice coefficient and IoU (Intersection over Union) to ensure the model maintained high accuracy in complex pulmonary embolism segmentation tasks.

[0099] Step 604: Further optimize the cross-validated model using the second loss function to obtain the final segmentation model.

[0100] Specifically, the second loss function includes the labeling loss function L. l and unlabeled loss function L u Two parts. The label loss function is calculated using the hard cross-entropy loss function H, which measures the difference between the model's prediction of the labeled image and the true label. Its expression is:

[0101]

[0102] Among them, B l Let H be the number of labeled images, and H be the hard cross-entropy loss function. Let be the predicted value for the i-th image. Let be the true label of the i-th image.

[0103] The unlabeled loss function achieves self-supervised learning through pseudo-labels generated by the model itself. It introduces diverse image enhancement methods (such as color enhancement, cropping, rotation, and scaling) and feature perturbation methods (such as Gaussian noise and feature pruning) to implement an enhancement perturbation strategy, thereby improving the model's robustness and adaptability to different pulmonary embolism morphologies. First, pseudo-labels are generated on weakly enhanced images, and then training is performed on strongly enhanced images. This weak-to-strong consistency regularization strategy maintains learning invariance on strongly enhanced images. The unlabeled loss function only updates the optimized segmentation model when the confidence of the pseudo-labels exceeds a preset confidence threshold, reducing the impact of noisy pseudo-labels. The expression for the unlabeled loss function is:

[0104]

[0105] Among them, B u Number of unlabeled images For the i-th pseudo-label of the weakly enhanced image, Let τ be the i-th pseudo-label of the strongly enhanced image, and τ be a preset confidence threshold. for The hard single-hot tags are obtained after the argmax operation.

[0106] The second loss function balances the contributions of the labeled and unlabeled loss functions to model training, and is expressed as: L = L l +λL u Where L is the second loss function, and λ is the weight coefficient. λ is usually set to 1, indicating that the labeled loss function and the unlabeled loss function are given the same weight.

[0107] It's worth noting that training the model using the second loss function effectively utilizes both limited labeled data and a large amount of unlabeled data, improving the model's performance in semi-supervised semantic segmentation tasks while reducing reliance on manual annotation and lowering costs. This allows the model to learn effectively from unlabeled data while maintaining high accuracy, thereby enhancing its generalization ability.

[0108] Step 700: Segment the recognition results using the final segmentation model to obtain the segmentation results;

[0109] Specifically, candidate regions containing pulmonary embolism in the identification results are selected as input to the final segmentation model. High-confidence pseudo-labels are generated based on the unlabeled parts in the input of the final segmentation model, and the identification results are segmented using the high-confidence pseudo-labels to obtain the segmentation results.

[0110] Step 800: Correct the segmentation result through the intermediary distance compensation mechanism to obtain the merged result.

[0111] Specifically, such as Figure 5 As shown, it includes:

[0112] Step 801: Calculate the intermediary distance between the segmentation results using a search algorithm.

[0113] Specifically, each detected thrombus boundary is extracted from the segmentation results. These boundaries are located outside the thrombus region. Using the vessel centerline as a reference, the location of the thrombus region within the vessel is determined. A search algorithm is used to calculate the intercalation distance between all detected thrombus boundaries, and a sorting algorithm is used to identify the nearest thrombi that meet the intercalation distance for each thrombus boundary. The intercalation distance refers to the pixel distance between the two closest thrombus boundary points along the vessel centerline in the segmentation results. The specific calculation method is as follows: find the two closest thrombus boundary points along the vessel centerline, and calculate the pixel distance between these two closest boundary points to obtain the intercalation distance. If the two thrombi are not on the same vessel, the distance between them along the vessel centerline is also calculated to obtain a global intercalation distance.

[0114] Step 802: Based on the comparison between the intercalation distance and the preset threshold, determine the two thrombi in the segmentation results to obtain the determination result. The determination result is: the same thrombus or different thrombi.

[0115] Specifically, the preset threshold is determined through experimental and empirical data. When the intercalation distance is less than the preset threshold, the two thrombi are judged as the same thrombus; otherwise, they are judged as different thrombi. In this embodiment, the preset threshold is 3 pixels.

[0116] Step 803: Merge the two thrombi that are determined to be the same thrombus to obtain the merged result.

[0117] Specifically, the boundary regions of two thrombi identified as belonging to the same thrombus are merged based on grayscale values ​​and their variations, forming a larger thrombus region. This merging operation is achieved by filling the space between the two thrombus regions with pixels. The region for filling pixels is determined through detailed analysis of the region's image grayscale features and variations. First, the mean, variance, and texture information of the thrombus regions to be merged are calculated to assess the region's consistency and naturalness. Subsequently, the clarity of the boundary and its separation from surrounding tissue are determined by evaluating the grayscale variations at the edges of the regions to be merged, such as gradients or grayscale change rates. Figures 6-8 As shown, the same thrombus was mistakenly identified as three thrombi, and the conclusion was that the three thrombi were the same thrombus. Figure 9 and Figure 10 These are the gaps between three thrombi, as shown in the three-dimensional diagram. Figure 11 As shown, Figures 12-14 All of these are results of thrombosis. A three-dimensional schematic diagram is shown below. Figure 15 As shown.

[0118] Furthermore, after merging, the morphology of the merged thrombus is checked for rationality to ensure that the spatial distribution of the new thrombus region corresponds to the actual anatomical structure. If the merged region is too large or irregular in shape, the morphology of the thrombus region is adjusted through further morphological manipulations (such as morphological expansion and erosion).

[0119] It should be noted that the merging strategy based on the intercalation distance of thrombi effectively eliminates multiple segmentations caused by model errors or low segmentation accuracy, thereby improving the accuracy of pulmonary embolism detection and reducing false alarms.

[0120] Furthermore, this embodiment also visualizes the merged results and generates a screening report. Volume rendering technology is used to transform CT images into intuitive 3D views. Combined with the detected thrombus location and confidence level, the pulmonary embolism features in the lung CT images are accurately displayed, allowing doctors to directly observe the 3D location distribution of pulmonary embolisms. Transparency, color coding, and texture variations are also used to visually distinguish thrombi and their confidence levels. Transparency is used to highlight thrombus areas of interest, while color coding differentiates based on the confidence level of the thrombus detection, facilitating doctors' rapid identification and assessment of the severity of pulmonary embolism. The screening report includes a 3D view of the patient's lung CT images, a confidence analysis of thrombus detection, segmentation results, and corresponding confidence levels, presented in a structured information format. This assists doctors in conducting comprehensive clinical diagnosis, significantly improving diagnostic efficiency.

[0121] The present invention also provides a pulmonary embolism detection and segmentation system based on intercalary distance compensation, comprising:

[0122] The image acquisition module is used to pre-annotate the lung CT image set to obtain pre-annotated labels;

[0123] The preprocessing module is used to perform preprocessing operations on the lung CT image set to obtain a standard image set;

[0124] The detection module integrates thrombus labels and standard image sets into a detection dataset. The detection dataset is then input into a 3D target detection network model for training, resulting in multiple detection training models. These multiple training models are then used to detect and identify the dataset, yielding multiple preliminary detection results. Finally, these preliminary detection results are fused together to obtain the final recognition result.

[0125] The segmentation module integrates thrombus labels, standard image sets, and lung CT image sets into a segmentation dataset. It then trains a semi-supervised semantic segmentation model using this dataset to obtain the final segmentation model. Finally, it segments the recognition results using the final segmentation model to obtain the segmentation result.

[0126] The accuracy compensation module is used to correct the segmentation results through the intercalation distance compensation mechanism to obtain the merged result;

[0127] The visualization and report generation module is used to visualize the merged results in three dimensions and generate screening reports.

[0128] The beneficial effects of this invention are as follows:

[0129] 1) By using computer artificial intelligence deep learning algorithms to detect and identify pulmonary thrombosis, the accuracy and reliability of detection have been significantly improved. It can effectively identify pulmonary embolism in complex clinical images and help doctors make diagnoses more quickly and accurately.

[0130] 2) Based on the morphological characteristics of different thrombi, multi-size anchor frames were designed, which improved the model's ability to detect thrombi of various shapes and sizes, ensuring comprehensive and accurate thrombus localization and effectively avoiding missed diagnoses and misdiagnoses.

[0131] 3) By using loss function optimization techniques, the accuracy of the model in locating the thrombus region was improved, and the overall robustness of the model was enhanced, so that consistent diagnostic results could be obtained in different types of cases.

[0132] 4) By adopting a semi-supervised learning method, which combines a small amount of labeled data with a large amount of unlabeled data, the value of unlabeled data is fully utilized, the performance and generalization ability of the segmentation model are improved, and the cost of manual annotation is reduced.

[0133] 5) The method of expanding the bounding box of the thrombus obtained from target detection by several pixels into the segmentation model is adopted. This method allows the segmentation model to focus on the contour boundary information around the small thrombus and obtain more accurate segmentation results.

[0134] 6) By introducing a merging strategy based on the intercalary distance of thrombi, the problem of missegmentation in the segmentation results is effectively solved, the phenomenon of multiple segmentation due to insufficient model accuracy is avoided, the accuracy of pulmonary embolism detection is improved, and the false alarm rate is reduced.

[0135] 7) By displaying the pulmonary embolism segmentation results through three-dimensional visualization, the spatial distribution of thrombi can be presented intuitively, helping doctors to better understand and analyze the characteristics of pulmonary embolism and assisting in clinical decision-making;

[0136] 8) Combining vascular centerline analysis provides a more accurate method for thrombus localization in pulmonary embolism detection, improves the accuracy of thrombus boundary, and ensures the precision of detection and segmentation;

[0137] 9) Tagging technology has improved the ability to identify pulmonary thrombosis and enhanced the accuracy of segmentation.

[0138] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0139] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A pulmonary embolism detection segmentation method based on intermedial distance compensation, characterized in that, The method comprises the following steps: pre-labeling a lung CT image set to obtain a pre-labeling label; the pre-labeling comprises manual labeling and pixel-level labeling; obtaining a thrombus label through the pre-labeling label; performing a preprocessing operation on the lung CT image set to obtain a standard image set; the preprocessing operation comprises adjusting a window width and window level, adjusting an image interval, adjusting an image direction cosine matrix, image denoising, and mixed cropping; integrating the thrombus label and the standard image set into a detection data set, inputting the detection data set into a three-dimensional target detection network model for model training to obtain a plurality of detection training models; detecting and recognizing the detection data set through the plurality of detection training models to obtain a plurality of preliminary detection results, and fusing the plurality of preliminary detection results to obtain a recognition result; integrating the thrombus label, the standard image set, and the lung CT image set into a segmentation data set, and performing model training on a semi-supervised semantic segmentation model through the segmentation data set to obtain a final segmentation model; segmenting the recognition result through the final segmentation model to obtain a segmentation result; modifying the segmentation result through an intermediate distance compensation mechanism to obtain a merging result; wherein integrating the thrombus label and the standard image set into a detection data set, inputting the detection data set into a three-dimensional target detection network model for model training to obtain a plurality of detection training models comprises: dividing the detection data set into a detection training set and a detection validation set; performing model training on the three-dimensional target detection network model through the detection training set to obtain a plurality of preliminary training models; extracting thrombus features through a multi-scale feature fusion technology on the detection training set; setting anchor boxes of multiple sizes according to the thrombus features, and determining thrombus regions through the anchor boxes; optimizing and training the preliminary training models through the thrombus regions and a first loss function to obtain the detection training models; modifying the segmentation result through an intermediate distance compensation mechanism to obtain a merging result comprises: calculating the intermediate distance between the segmentation results through a search algorithm; judging two thrombi in the segmentation result according to a comparison result of the intermediate distance and a preset threshold to obtain a judgment result; the judgment result is the same thrombus or different thrombi; merging two thrombi with the same thrombus in the judgment result to obtain a merging result; calculating the intermediate distance between the segmentation results through a search algorithm, specifically: extracting thrombus boundaries from the segmentation results, and calculating the pixel distance between the closest points of two thrombi through a blood vessel center line and the thrombus boundaries to obtain the intermediate distance.

2. The method of claim 1, wherein, obtaining a thrombus label through a pre-labeling label, specifically: performing coordinate detection on the lung CT image according to the pre-labeling label to obtain a thrombus bounding box, and obtaining the thrombus label according to the 6 corner coordinate scalars of the thrombus bounding box; the thrombus bounding box is a cuboid structure.

3. The method of claim 1, wherein, The first loss function comprises a position loss and a confidence loss; a calculation formula of the first loss function is: ; wherein, GIoU is the first loss function, IoU is an IoU loss function, A is an optimal anchor box, B is a thrombus label, C is a minimum closed box formed by A and B, |C| is an area of the minimum closed box, and |C\(A∪B)| is an area of C after removing the union set of A and B.

4. The method of claim 1, wherein, The thrombus label, the standard image set and the lung CT image set are integrated into a segmentation data set, a semi-supervised semantic segmentation model is trained through the segmentation data set, and a final segmentation model is obtained. Specifically, the segmentation data set is input into the semi-supervised semantic segmentation model, and the final segmentation model is obtained through consistency regularization loss, enhanced disturbance strategy, pseudo label generation strategy, cross validation and second loss function.

5. The method of claim 4, wherein, An expression of the second loss function is: L = L l + λL u ; wherein, L is the second loss function value, L l is a labeled loss function, L u is an unlabeled loss function, and λ is a weight coefficient. The expression of the mark loss function is: ; wherein, B l is the number of mark images, H is a hard cross-entropy loss function, is the predicted value of the i th image, is the true label of the i th image; The expression of the unmarked loss function is: ; wherein B u is the number of unmarked images, is the i-th pseudo label of the weak augmented image, is the i-th pseudo label of the strong augmented image, τ is a preset confidence threshold, is a hard one-hot label.

6. The inter-distance compensation based pulmonary embolism detection segmentation method of claim 1, wherein, The final segmentation model is used to segment the recognition result, and a segmentation result is obtained. Specifically, a candidate region containing a pulmonary embolism in the recognition result is selected as the input of the final segmentation model, high-confidence pseudo labels are generated according to the unannotated part in the input of the final segmentation model, and the recognition result is segmented through the high-confidence pseudo labels to obtain the segmentation result.

7. A pulmonary embolism detection segmentation system based on intermediate distance compensation for use in the segmentation method of any one of claims 1-6, characterized in that, Comprise: An image acquisition module is configured to pre-annotate a lung CT image set to obtain pre-annotation labels. A preprocessing module is configured to perform a preprocessing operation on the lung CT image set to obtain a standard image set. A detection module is configured to integrate the thrombus label and the standard image set into a detection data set, input the detection data set into a three-dimensional target detection network model for model training, and obtain a plurality of detection training models. The detection data set is detected through a plurality of detection training models to obtain a plurality of preliminary detection results, and the preliminary detection results are fused to obtain a recognition result. A segmentation module is configured to integrate the thrombus label, the standard image set and the lung CT image set into a segmentation data set, train a semi-supervised semantic segmentation model through the segmentation data set, and obtain a final segmentation model. The final segmentation model is used to segment the recognition result, and a segmentation result is obtained. An accuracy compensation module is configured to correct the segmentation result through an intermediate distance compensation mechanism to obtain a merged result. A visualization and report generation module is configured to perform three-dimensional visualization on the merged result and generate a screening report.

Images