A method for identifying pulmonary embolism based on non-contrast CT

Abstract

The present invention provides a method for identifying pulmonary embolism based on non-contrast CT, belonging to the technical field of deep learning. In order to solve the problem in the prior art that the identification of pulmonary embolism mostly relies on contrast-enhanced CT and it is impossible to quickly and accurately identify pulmonary embolism for non-contrast CT data. The method of the present invention performs pulmonary contour segmentation on non-contrast CT images through a pulmonary contour segmentation model; performs segmentation of pulmonary blood vessels in non-contrast CT through a blood vessel segmentation model; through a pulmonary embolism extraction model, uses the pulmonary embolism area on contrast-enhanced CT as the gold standard to mark the pulmonary embolism area in non-contrast CT, determines the embolism suspicious area by considering the CT value difference between pulmonary embolism and blood vessels and the vascular morphological features caused by pulmonary embolism for the pulmonary embolism in non-contrast CT, optimizes the model by minimizing the difference between the suspicious area and the gold standard, and uses the optimized model to perform pulmonary embolism segmentation on non-contrast CT. The present invention can achieve high sensitivity and accuracy in predicting pulmonary embolism based on non-contrast CT.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning, and in particular, to a method for identifying pulmonary embolism based on non-contrast CT. Background Art

[0002] Pulmonary embolism (PE) is a clinical and pathophysiological syndrome in which emboli from inside or outside the body block the pulmonary artery, causing disorders of the pulmonary circulation. Among them, 70%-95% is caused by the detachment of thrombi in the lower extremities and pelvis after a person has been sedentary for a long time, and then enters the pulmonary artery and its branches through the blood circulation, resulting in blockage. Therefore, pulmonary embolism only occurs inside the pulmonary artery, and has the characteristics of low diagnostic rate and high mortality. Due to the different sizes of the emboli, the clinical manifestations of patients are also different. When the emboli are small, patients have almost no symptoms or show mild chest tightness. If the area of the blood vessel blocked by the emboli is large, patients may experience fainting or even sudden death. In recent years, pulmonary embolism has become a disease with a relatively high incidence and mortality in China.

[0003] Common examination methods for pulmonary embolism, such as D-dimer, blood gas analysis, ultrasound, etc., cannot confirm the diagnosis due to poor specificity. At present, the main method for diagnosing PE is pulmonary CT angiography (CTPA). The realization of this technology is that a doctor injects a contrast agent into the patient's blood vessel, and then performs a CT scan. The contrast agent makes the pulmonary blood vessels image more clearly under the CT scanner, and the imaging effect is better than that of ordinary non-contrast CT. At present, CT pulmonary angiography (CTPA) is the gold standard for diagnosing pulmonary embolism, but CTPA has its disadvantages, such as radiation damage, radiation exposure, contrast agent-related nephropathy, and high cost, etc. It is not suitable for repeated examinations in a short period of time and is difficult to be widely used clinically. Therefore, there is an urgent need for a method for identifying pulmonary embolism based on non-contrast CT clinically, so that the use of contrast agents is more targeted, thereby improving the treatment effect of pulmonary embolism, reducing the hospitalization time of patients, reducing the incidence of pulmonary embolism, and at the same time avoiding the economic losses caused by excessive use of contrast agents and the negative impacts on the patient's body. Summary of the Invention

[0004] The technical problem to be solved by the present invention is as follows:

[0005] In the prior art, the identification of pulmonary embolism lesions mostly relies on enhanced CT that requires injection of contrast agents, and it is impossible to quickly and accurately identify pulmonary embolism for non-contrast CT data.

[0006] The technical solution adopted by the present invention to solve the above technical problem is as follows:

[0007] The present invention constructs a method for identifying pulmonary embolism based on non-contrast CT data for non-contrast CT data, uses the pulmonary embolism area on enhanced CT as the gold standard to construct a pulmonary embolism extraction model, improves the extraction accuracy of the model for pulmonary embolism, and reduces errors, including the following steps:

[0008] Step 1: Collect non-contrast CT data of the lungs of patients with pulmonary embolism, perform quality control on the data, remove the "risky" data to obtain high-quality non-contrast CT data, and standardize the data;

[0009] Step 2: Construct a lung contour segmentation model, perform segmentation processing on the standardized non-contrast CT images, remove the regions unrelated to the lungs, and obtain a lung contour image containing the effective lung region;

[0010] Step 3: Construct a reconstruction model to reconstruct the thick-slice non-contrast CT image into a thin-slice non-contrast CT image, construct a blood vessel segmentation model to segment the pulmonary blood vessels of the thin-slice non-contrast CT, and simultaneously input the lung contour image obtained in Step 2 to improve the blood vessel segmentation effect and efficiency of the model, and obtain a pulmonary artery image;

[0011] Step 4: Construct a pulmonary embolism extraction model, use the standardized non-contrast CT image as the input, and input the pulmonary artery image obtained in Step 3 to guide the model to focus on the blood vessel region; collect the enhanced CT data of the lungs corresponding to the standardized CT data, and have professional physicians mark the pulmonary embolism regions in the enhanced CT data of the lungs, and use the pulmonary embolism regions on the enhanced CT as the gold standard; the pulmonary embolism extraction model determines the suspicious embolism regions by considering the CT value differences between pulmonary embolism and blood vessels and the vascular morphological features caused by pulmonary embolism, compares the obtained suspicious regions with the gold standard of pulmonary embolism in the enhanced CT, constructs a loss function by minimizing the difference between the two, and optimizes the model through the backpropagation algorithm to obtain the pulmonary embolism extraction model;

[0012] Step 5: Use the obtained pulmonary embolism extraction model to segment the pulmonary embolism in the non-contrast CT.

[0013] Further, the quality control in Step 1 includes the following steps:

[0014] Step 1-1: Use a preprocessing segmentation model to remove the regions unrelated to the lungs to obtain a lung parenchyma image;

[0015] Step 1-2: Perform 3D visualization on the obtained lung parenchyma image, and mark the CT data quality with obvious defects in the segmentation result as "risky";

[0016] Step 1-3: Perform a "smoothness" analysis on the obtained lung parenchyma segmentation image. If the smoothness exceeds the set range, mark the CT data quality as "risky";

[0017] Step 14: Layer-by-layer inspection is performed on the marked "risky" CT data to determine whether the CT data quality is poor or there are errors in the preprocessing segmentation model. If there are errors in the model, the model structure or parameters are modified in a timely manner.

[0018] Furthermore, the lung contour segmentation model is based on a 2.5D segmentation algorithm and uses a weighted attention mechanism, dilated convolution, and cross-entropy loss.

[0019] Furthermore, the reconstruction model in Step 3 is based on the ResU-net network and is composed of multiple groups of non-linear SR feature blocks connected in series. The feature blocks are composed of 3×3 convolution (Conv) kernels, instance normalization, and Leaky ReLU activation functions, and two intersecting attention modules are inserted between the non-linear feature blocks.

[0020] Furthermore, the blood vessel segmentation model is a two-stage 2.5D segmentation model using feature enhancement loss. A high-recall blood vessel image and a high-precision blood vessel image are output by the first-stage segmentation model. The obtained high-recall blood vessel image, high-precision blood vessel image, and normalized CT data are input into the second-stage segmentation model, and finally the pulmonary artery segmentation result is obtained.

[0021] Furthermore, each 2.5D segmentation model of the blood vessel segmentation model uses a greedy algorithm to find an approximate optimal solution, and optimized parameters are set for the 2.5D model.

[0022] Furthermore, the pulmonary embolism extraction model is based on the transformer network. In the encoding stage of the transformer, CNN and DNN are used to extract low-level features, and then the global interaction is modeled by the Transformer to capture high-level semantic context. The pulmonary artery segmentation result is embedded in the positional encoding part, so that the model only focuses on the internal area of the pulmonary artery; in the decoding stage, a guidance module is established to guide the segmentation of the pulmonary embolism. After the blood vessel encoding information is input into the guidance module, the pulmonary artery segmentation result is used as the positional embedding and embedded into the multi-head attention mechanism, and then after normalization and concat operations, it is used as the output of the guidance module. Finally, the pulmonary embolism extraction result is obtained through softmax and reshape operations.

[0023] Furthermore, the loss function of the pulmonary embolism extraction model uses the cross-entropy loss function, and the formula is:

[0024] weighted_loss=-w fn ×ln(p)×p′-w fp ln(1 - p)×(1 - p′)

[0025] Among them, p is the predicted probability that the voxel is positive, and p′ is the true annotation (ground truth) probability that the voxel is positive. Through w fn Adjust the false negative penalty. Through w fp Adjust the false positive penalty. When the ground truth is binary, if the voxel is positive (p′ equals 1), a w fn needs to be defined. If the voxel is negative ((1 - p′) equals 1), a w fp needs to be defined.

[0026] A system for quickly and accurately identifying pulmonary embolism based on non-contrast CT. The system has program modules corresponding to the steps of any of the above technical solutions, and when running, it executes the steps in the above method for identifying pulmonary embolism based on non-contrast CT.

[0027] A computer-readable storage medium stores a computer program configured to implement the steps of the method for identifying pulmonary embolism based on non-contrast CT according to any of the above technical solutions when called by a processor.

[0028] Compared with the prior art, the beneficial effects of the present invention are:

[0029] In the method for identifying pulmonary embolism based on non-contrast CT of the present invention, during the process, the lung contour extraction model, the blood vessel segmentation model, and the pulmonary embolism extraction model are respectively used to segment and identify the lung contour, the pulmonary blood vessels, and the pulmonary embolism. For non-contrast CT data, the transformer model is used to extract pulmonary embolism, and the embolism suspicious area is determined by combining the difference in the average CT value between the pulmonary embolism and the blood vessel, the morphological characteristics of the blood vessel caused by the pulmonary embolism, and the blood vessel segmentation result. The embolism suspicious area is compared with the gold standard of pulmonary embolism in enhanced CT, so that the pulmonary embolism extraction model has the ability to extract pulmonary embolism, and can achieve high sensitivity and accuracy in predicting pulmonary embolism based on non-contrast CT.

[0030] In the method of the present invention, using the lung contour image as a guide significantly improves the segmentation effect and efficiency of arteries. Then, using the artery image as a guide enables the pulmonary embolism extraction model to have the ability to extract pulmonary embolism, thereby achieving the purpose of identifying pulmonary embolism in non-contrast CT data through the pulmonary embolism extraction model, and the model has high performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is the flowchart of the method for identifying pulmonary embolism based on non-contrast CT in the embodiment of the present invention;

[0032] Figure 2 is the flowchart of CT data spatial normalization in the embodiment of the present invention;

[0033] Figure 3This is the result diagram during the process of segmenting blood vessels by the blood vessel segmentation model in the example of the present invention;

[0034] Figure 4 This is the flowchart of the 2.5D segmentation algorithm in the example of the present invention;

[0035] Figure 5 This is the segmentation result diagram of pulmonary embolism in the example of the present invention;

[0036] Figure 6 This is the comparison diagram between the pulmonary embolism area predicted by the pulmonary embolism extraction model in the plain CT and the gold standard of the pulmonary embolism area in the enhanced CT in the example of the present invention. Detailed implementation manners

[0037] In the description of the present invention, it should be noted that in the embodiments of the present invention, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features.

[0038] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following detailed description of the specific embodiments of the present invention will be given with reference to the accompanying drawings.

[0039] Combined with Figures 1 to 6 As shown, first, CT scan data of 1826 pulmonary embolism patients from five hospitals are collected (including both plain CT and enhanced CT, and the time interval between the two examinations is less than two days to ensure that the size and location of the embolism do not change significantly during the interval). The scan data is obtained using a Philips iCT 256 or UIH uCT 528 commercial CT scanner.

[0040] Perform quality control on the data, and use a preprocessing segmentation model, that is, a model based on the 2.5D segmentation algorithm, including a coronal and sagittal feature extraction network and a feature weighted fusion attention module. According to the semantic features of the lungs, remove the regions irrelevant to the lungs to obtain the lung parenchyma image;

[0041] Perform 3D visualization on the obtained lung parenchyma image by stereolithography, and mark the CT data with obvious defects in the segmentation result as "at risk";

[0042] Perform a "smoothness" analysis on the obtained lung parenchyma segmentation image. If the smoothness exceeds the set range [17.11, 23.03], mark the CT data quality as "at risk";

[0043] For the marked "risky" CT data (about 5%), layer-by-layer inspection is carried out to determine whether it is due to poor CT data quality or errors in the preprocessing segmentation model. If it is an error in the model, the model structure or parameters are modified in a timely manner to remove the "risky" data and obtain high-quality plain CT data.

[0044] As Figure 2 shown, data standardization is performed on the obtained high-quality plain CT data. Data standardization includes spatial normalization and signal normalization. First, the data is spatially normalized, and at the same time, the voxel resolution and data dimension are normalized; after normalization, the volume corresponding to each voxel is such that where δ represents the volume of each voxel of the CT data, and in this embodiment, its value is 334×334×512mm 3 , and spatial normalization does not lose data information.

[0045] The formula used for signal normalization is:

[0046]

[0047] where, I original is the original CT value, WL is the window level, and WW is the window width.

[0048] A lung contour segmentation model is constructed to perform segmentation processing on the plain CT image, remove the regions irrelevant to the lungs, and generate a lung contour image containing the effective region of the lungs.

[0049] The standardized CT data is randomly divided into a training set (70%), a test set (20%), and a validation set (10%). The labelme software is used to label the lung regions in the training set to convert the image into a binary image.

[0050] The lung contour segmentation model is based on a 2.5D segmentation algorithm and uses a weighted attention mechanism, dilated convolution, and cross-entropy loss.

[0051] The accuracy and reliability of lung contour segmentation depend on the diversity of the training data. In the present invention, a large amount of multi-center lung CT data is collected. Since pulmonary embolism lesions include some common abnormal diseases such as: pulmonary fibrosis, pneumothorax, masses, etc. When using the labelme software to label the lung regions, the above common lesions are also retained, providing favorable conditions for the final identification of pulmonary embolism.

[0052] As Figure 4As shown, in the 2.5D segmentation algorithm of the present invention, a two-dimensional U-net is used to perform two-dimensional segmentation on the XY, YZ, and XZ directional planes, and ensemble learning is used to combine the two-dimensional segmentation results from different perspectives to output the final three-dimensional segmentation result; the input of the model is a standardized machine-independent tensor with a shape of The inputs in the three directions have the same shape and dimensions, both stacked by m adjacent CT slices and n guiding channels. The output of the 2D model is a probability map with a shape of indicating the probability that the voxel is the lung in that direction. Through fusing the probability maps, a 3D segmentation model with a shape of is obtained.

[0053] A reconstruction model is constructed to reconstruct thick-slice plain CT images into thin-slice plain CT images. A vascular segmentation model is constructed to segment the pulmonary blood vessels in the thin-slice plain CT, and the lung contour image obtained in the second step is input simultaneously to improve the vascular segmentation effect and efficiency of the model, and a pulmonary artery image is obtained.

[0054] The reconstruction model is based on the ResU-net network to restore the spatial continuity of features on adjacent slices. The model is composed of 14 groups of non-linear SR feature blocks connected in series, and these feature blocks are composed of 3×3 convolutional (Conv) kernels, instance normalization, and LeakyReLU; two intersecting attention modules are between the 2nd and 3rd groups, and the 13th and 14th groups. Thick-slice CT (slice thickness > 2.5mm) can be reconstructed into thin-slice plain CT (inter-slice thickness <= 1.00mm). Since vascular features are not easily extracted in plain CT, the thick-slice plain CT images can be reconstructed into thin-slice plain CT images through the reconstruction model to accurately reconstruct the vascular features required for artery-vein segmentation.

[0055] As Figure 3 shown, the vascular segmentation model is a two-stage 2.5D segmentation model using feature-enhanced loss. Through the first-stage segmentation model, a high-recall vascular image (setting the probability threshold for voxels being blood vessels to about 0.5) and a high-precision vascular image (setting the probability threshold for voxels predicted as blood vessels to about 0.8) are output. The obtained high-recall vascular image, high-precision vascular image, and standardized CT data are input into the second-stage segmentation model. The high-recall vascular image and high-precision vascular image provide precise guidance for the second-stage segmentation model, enabling the model to focus on and refine small regions, and at the same time reducing the model's search space by thousands of times.

[0056] A training set (70%), a test set (20%), and a validation set (10%) of segmented standardized CT data are adopted. Since pulmonary embolism exists in the pulmonary artery, but there are certain differences between the pulmonary artery and the pulmonary vein, the pulmonary artery and the pulmonary vein are respectively labeled to improve the segmentation accuracy. The labelme software is used to mark the regions of the pulmonary artery and the pulmonary vein in the training set, so that the image is converted into a binary image.

[0057] The training set data is used to train the vascular segmentation model, the test set data is used to test the performance of the vascular segmentation model, and the trained vascular segmentation model is used to segment the CT data of the validation set. Moreover, during the training process of the vascular segmentation model, a fluctuating training protocol is adopted, which changes the global class balance weight of the loss function, thereby changing the relative penalty between class mispredictions and avoiding the model's performance from falling into local minima.

[0058] Each 2.5D segmentation model of the vascular segmentation model adopts a greedy algorithm to obtain an approximate optimal solution, and optimized parameters are set for the 2.5D model. The problem is divided into four sub-problems: loss function, adjacent slices, training protocol, and initial features. For each 2.5D model, up to 12 parameters need to be evaluated. Compared with the default parameters of the 2.5D model, the optimized parameters can increase the dice value by about 0.1.

[0059] In the cross-sectional image of the lungs input to the vascular segmentation model, the images of the blood vessels appear as discontinuous regions. The cross-section of the aorta in the region is hundreds of pixels, while the cross-section of small blood vessels is only a few pixels. Traditional loss functions are based on voxel-level performance (such as voxel-level cross-entropy loss, dice loss, etc.) and pay less attention to tiny structure regions because the total area of all tiny structure regions is much smaller than that of large regions, which will lead to misdetection of tiny structures. Therefore, the present invention proposes a feature enhancement loss, which helps the model extract tiny structures. The feature enhancement loss is a voxel-level balanced cross-entropy loss, which is the sum of all voxel losses. For each voxel, the loss is defined as:

[0060] voxelloss = -w × ln(p) × p' - ln(1 - p) × (1 - p')

[0061] where p is the predicted probability that the voxel is positive, p' is the true probability that the voxel is positive, and w is the penalty weight for prediction errors.

[0062] Construct a pulmonary embolism extraction model, which takes non-contrast CT images as input and inputs the pulmonary artery images to guide the model to focus on the vascular region; collect contrast-enhanced CT data of the lungs corresponding to the standardized CT data, and mark the pulmonary embolism regions in the contrast-enhanced CT data of the lungs by professional physicians, and use the pulmonary embolism regions on the contrast-enhanced CT as the gold standard; the pulmonary embolism extraction model determines the suspicious embolism regions by considering the CT value differences between pulmonary embolisms and blood vessels, where the regions with abnormal CT values appear as multiple, continuous scattered points similar to noise, and also considering the vascular morphological features caused by pulmonary embolisms. Then, compare the CT values and vascular morphology of the obtained suspicious regions with the pulmonary embolism gold standard in the contrast-enhanced CT, construct a loss function by minimizing the difference between the two, and optimize the model through the backpropagation algorithm. After iterative calculation, the pulmonary embolism extraction model is obtained.

[0063] The pulmonary embolism extraction model is based on the Transformer network. In the encoding stage of the Transformer, CNN and DNN are used to extract low-level features, and then the Transformer is used to model global interactions to capture high-level semantic contexts. The pulmonary artery segmentation results are embedded in the positional encoding part, so that the model only focuses on the internal region of the pulmonary artery; in the decoding stage, a guiding module is established to guide the segmentation of pulmonary embolisms. When the vascular encoding information is input into the guiding module, the pulmonary artery segmentation results are used as positional embeddings and embedded into the multi-head attention mechanism, and then, after normalization and concat operations, they are used as the output of the guiding module. Finally, the pulmonary embolism extraction results are obtained through softmax and reshape operations.

[0064] Use the training set (70%), test set (20%), and validation set (10%) of the divided standardized CT data. Use the training set data to train the pulmonary embolism extraction model, use the test set data to test the performance of the pulmonary embolism extraction model, and use the trained pulmonary embolism extraction model to segment the validation set CT data.

[0065] The inventor found through research that since the density of the thrombus site is higher than that of the normal blood vessels, its average CT value is higher than the average CT value inside the pulmonary artery. Through measurement, it is found that the average CT of the thrombus position is 20-100 HU higher than the CT value of the same-level blood vessel branches, and the value at the center of the thrombus is significantly higher than that on the surface of the thrombus; in terms of vascular morphology, the embolism site presents a "vascular cast" or truncation, and the diameter of the blocked segment of the blood vessel can be thickened or its morphology can be changed to varying degrees.

[0066] The loss function of the pulmonary embolism extraction model uses the cross-entropy loss function, and on this basis, voxel weights are added to balance the problem that small embolisms are ignored. By weighting, the penalties of false positives and false negatives are changed. The formula of the loss function is:

[0067] weighted_loss = -w fn ×ln(p)×p′ - w fp ln(1 - p)×(1 - p′)

[0068] Where p is the predicted probability that the voxel is positive, and p′ is the true annotation (ground truth) probability that the voxel is positive. The false negative penalty is adjusted by w fn and the false positive penalty is adjusted by w fp When the ground truth is binary, if the voxel is positive (p′ equals 1), a w needs to be defined fn , and if the voxel is negative ((1 - p′) equals 1), a w needs to be defined fp .

[0069] To obtain the final pulmonary embolism extraction model, the non - enhanced CT is standardized. The pulmonary embolism in the standardized non - enhanced CT is segmented using the obtained pulmonary embolism extraction model, and the pulmonary embolism extraction result as shown Figure 5 is obtained.

[0070] The Dice value for pulmonary artery model segmentation by the method of the present invention can reach 0.919. Compared with existing methods such as machine learning models, the method of the present invention has high accuracy in segmenting the pulmonary artery from non - enhanced CT.

[0071] As shown Figure 6 , the results of the pulmonary embolism extraction model of the present invention for predicting the pulmonary embolism region in non - enhanced CT are relatively consistent with those in enhanced CT, indicating that the method of the present invention has high accuracy in segmenting pulmonary embolism from non - enhanced CT. The segmentation effect of the pulmonary embolism extraction model of the present invention is evaluated by AUROC (Area Under the Receiver Operating Characteristic curve), sensitivity, and specificity: the AUROC is 0.85, the sensitivity and specificity are 76.0% and 80.5% respectively. At the same time, the model can detect very subtle abnormal increases in values. When the average increase in thrombus is 30 HU, the Dice value for the model to detect the abnormal increase can reach greater than 0.75.

[0072] Although the present invention is disclosed as above, the protection scope of the present invention is not limited thereto. Those skilled in the art of the present invention can make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will all fall within the protection scope of the present invention.

Claims

1. A method for identifying pulmonary embolism based on non-contrast CT, characterized in that It includes the following steps: Step 1: Collect the plain chest CT data of patients with pulmonary embolism, conduct quality control on the data, remove the "risky" data, obtain high-quality plain chest CT data, and standardize the data; Step 2: Build a lung contour segmentation model, perform segmentation processing on the standardized plain chest CT images, remove the regions unrelated to the lungs, and obtain a lung contour image containing the effective lung regions; Step 3: Build a reconstruction model to reconstruct the thick-slice plain chest CT images into thin-slice plain chest CT images, build a blood vessel segmentation model to segment the pulmonary blood vessels in the thin-slice plain chest CT, and simultaneously input the lung contour image obtained in Step 2 to improve the blood vessel segmentation effect and efficiency of the model, and obtain a pulmonary artery image; Step 4: Build a pulmonary embolism extraction model, use the standardized plain chest CT images as input, and input the pulmonary artery image obtained in Step 3 to guide the model to focus on the blood vessel regions; collect the enhanced chest CT data corresponding to the standardized CT data, and have professional physicians mark the pulmonary embolism regions in the enhanced chest CT data, and use the pulmonary embolism regions on the enhanced CT as the gold standard; the pulmonary embolism extraction model determines the suspicious embolism regions by considering the CT value differences between pulmonary embolisms and blood vessels and the vascular morphological features caused by pulmonary embolisms, compares the obtained suspicious regions with the pulmonary embolism gold standard in the enhanced CT, constructs a loss function by minimizing the differences between the two, and optimizes the model through the backpropagation algorithm to obtain the pulmonary embolism extraction model; Step 5: Use the obtained pulmonary embolism extraction model to segment the pulmonary embolisms in the plain chest CT; The pulmonary embolism extraction model is based on the transformer network. In the encoding stage of the transformer, CNN and DNN are used to extract low-level features, and then the global interaction is modeled by the Transformer to capture high-level semantic contexts. The pulmonary artery segmentation results are embedded in the positional encoding part, so that the model only focuses on the internal regions of the pulmonary artery; in the decoding stage, a guiding module is established to guide the segmentation of pulmonary embolisms. When the blood vessel encoding information is input into the guiding module, the pulmonary artery segmentation results are used as positional embeddings and embedded into the multi-head attention mechanism, and then, after normalization and concat operations, they are used as the output of the guiding module. Finally, the pulmonary embolism extraction results are obtained through softmax and reshape operations; The loss function of the pulmonary embolism extraction model uses the cross-entropy loss function, and the formula is: weighted_loss = -w fn ×ln(p)×p′ - w fp ln(1 - p)×(1 - p′) where p is the predicted probability that the voxel is positive, p' is the true annotation probability that the voxel is positive, and the false negative penalty is adjusted by w fn The false positive penalty is adjusted by w fp When the true annotation is binary, if the voxel is positive, a w needs to be defined fn and if the voxel is negative, a w needs to be defined fp .

2. The method for identifying pulmonary embolism based on non-contrast CT according to claim 1, wherein The quality control in Step 1 includes the following steps: Step 1-1: Use a preprocessing segmentation model to remove the regions unrelated to the lungs to obtain a lung parenchyma image; Step 1-2: Perform 3D visualization on the obtained lung parenchyma image, and mark the CT data quality with obvious defects in the segmentation results as "risky"; Step 1-3: Conduct a "smoothness" analysis on the obtained lung parenchyma segmentation image. If the smoothness exceeds the set range, mark the CT data quality as "risky"; Step Fourteen: Layer-by-layer inspection is performed on the marked "at-risk" CT data to determine whether the CT data quality is poor or there are errors in the preprocessing segmentation model. If there are errors in the model, the model structure or parameters are modified in a timely manner.

3. The method for identifying pulmonary embolism based on non-contrast CT according to claim 1, wherein The lung contour segmentation model is based on a 2.5D segmentation algorithm and uses a weighted attention mechanism, dilated convolution, and cross-entropy loss.

4. The method for identifying pulmonary embolism based on non-contrast CT according to claim 1, wherein The reconstruction model in Step Three is based on the ResU-net network and consists of multiple groups of non-linear SR feature blocks connected in series. The feature blocks are composed of a 3×3 convolution kernel, instance normalization, and a Leaky ReLU activation function, and two intersecting attention modules are inserted between the non-linear feature blocks.

5. The method for identifying pulmonary embolism based on non-contrast CT according to claim 4, characterized in that The vascular segmentation model is a two-stage 2.5D segmentation model using feature enhancement loss. A high-recall vascular image and a high-precision vascular image are output by the first-stage segmentation model. The obtained high-recall vascular image, high-precision vascular image, and normalized CT data are input into the second-stage segmentation model, and finally the pulmonary artery segmentation result is obtained.

6. The method for identifying pulmonary embolism based on non-contrast CT according to claim 5, wherein Each 2.5D segmentation model of the vascular segmentation model uses a greedy algorithm to find an approximate optimal solution, and optimized parameters are set for the 2.5D model.

7. A fast and accurate pulmonary embolism recognition system based on non-contrast CT, characterized in that The system has program modules corresponding to the steps of any one of the above claims 1 to 6, and when running, executes the steps in the above method for identifying pulmonary embolism based on non-contrast-enhanced CT.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of the method for identifying pulmonary embolism based on non-contrast-enhanced CT according to any one of claims 1 to 6 when called by a processor.

Images