Intelligent risk layering method and system for pulmonary embolism and storage medium
By using a neural network model based on the U-Net architecture and image omics feature analysis, the automatic segmentation and risk stratification of pulmonary embolism lesions were integrated, which solved the problems of operational complexity and insufficient generalization ability caused by the independent segmentation and prediction steps in the existing technology, and improved diagnostic efficiency and accuracy.
Patent Information
- Application Number
- CN202510806081.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-17
- Publication Date
- 2025-12-09
AI Technical Summary
In existing technologies, when using artificial intelligence models to predict the risk stratification of pulmonary embolism, the image segmentation and risk stratification prediction steps are independent and complex, making it difficult to balance accuracy and generalization ability, resulting in diagnostic efficiency that does not meet clinical needs.
A neural network model based on the U-Net architecture is used for image segmentation, and a predictive model based on radiomics feature analysis is combined to achieve automatic segmentation and risk stratification of pulmonary embolism lesions. By training and adjusting the model parameters, the segmentation accuracy and prediction accuracy are improved.
It enables efficient automatic segmentation and risk stratification of pulmonary embolism lesions, improving the overall quality and accuracy of diagnosis and meeting clinical needs.
Smart Images
Figure CN121095133A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of automated risk prediction technology for diseases, and in particular to a method and system for risk stratification of pulmonary embolism, and a storage medium. Background Technology
[0002] CTPA (Computed Tomography Pulmonary Angiography) is a non-invasive imaging examination that uses CT technology to evaluate the pulmonary arteries and their branches. It provides a comprehensive assessment of the pulmonary vascular system and is mainly used to diagnose pulmonary embolism. Doctors can also use CTPA images to stratify and predict the risk of pulmonary embolism in patients.
[0003] Previously, the prediction of pulmonary embolism risk stratification relied mainly on the professional experience of physicians. However, different physicians may interpret the same CTPA image differently, potentially leading to misjudgments in risk stratification. To address this, the industry has proposed a technique that uses artificial intelligence models to analyze CTPA images to derive the pulmonary embolism risk stratification structure.
[0004] However, current methods for predicting the risk of pulmonary embolism using artificial intelligence models are often divided into two completely independent steps: image segmentation and risk stratification prediction. This leads to operational complexity and interruption issues, a lack of overall integration, and difficulty in balancing the accuracy and generalization ability of different models, resulting in diagnostic efficiency that fails to meet clinical needs. Summary of the Invention
[0005] In view of this, this application provides an integrated method and system for risk stratification of pulmonary embolism lesions, as well as a storage medium.
[0006] Firstly, a risk stratification method for pulmonary embolism is proposed, including:
[0007] The CTPA image is input into the image segmentation model to obtain a segmented image with the lesion region marked by the image segmentation model.
[0008] The segmented image is input into the prediction model to obtain the pulmonary embolism risk stratification prediction result output by the prediction model, wherein the pulmonary embolism risk stratification prediction result is determined by the prediction model based on the radiomics features of the lesion region, and the radiomics features are extracted by the prediction model from the lesion region.
[0009] In some possible implementations, the radiomics features include embolus volume, CT value, and clot burden score.
[0010] In some possible implementations, before inputting the CTPA image into the image segmentation model, the method further includes training the image segmentation model, specifically including:
[0011] Obtain m first CTPA images and m second CTPA images with the same image content as the m first CTPA images, wherein each first CTPA image is additionally marked with a first lesion region relative to the corresponding second CTPA image, and m is an integer not less than 100;
[0012] Input n of the m second CTPA images into the image segmentation model to obtain n first segmented images with the second lesion region marked by the image segmentation model, where n < m;
[0013] If the difference between the second lesion region of the n first segmented images and the first lesion region of the corresponding n first CTPA images is greater than a first difference threshold, then the parameters of the image segmentation model are adjusted until the difference between the second lesion region of the n segmented images and the first lesion region of the corresponding n first CTPA images is less than the first difference threshold, and thereafter...
[0014] The mn second CTPA images, excluding the n second CTPA images, are input into the image segmentation model to obtain mn second segmented images with the third lesion region marked by the image segmentation model.
[0015] If the difference between the third lesion region of the mn second segmented images and the first lesion region of the corresponding mn first CTPA images is greater than a second difference threshold, then the parameters of the image segmentation model are adjusted until the difference between the second lesion region of the n segmented images and the first lesion region of the corresponding n first CTPA images is less than the first difference threshold, and the difference between the third lesion region of the mn second segmented images and the first lesion region of the corresponding mn first CTPA images is less than the second difference threshold, wherein the third difference threshold is smaller than the first difference threshold.
[0016] In some possible implementations, the second difference threshold is equal to the first difference threshold, and the image segmentation model is a neural network model based on the U-net architecture.
[0017] In some possible implementations, the step of the difference between the second lesion region of the n segmented images and the first lesion region of the corresponding n first CTPA images being less than a first difference threshold, and the difference between the third lesion region of the mn second segmented images and the first lesion region of the corresponding mn first CTPA images being less than a second difference threshold, includes:
[0018] The difference between the second lesion region of the n segmented images and the first lesion region of the corresponding n first CTPA images is less than a third difference threshold, and the difference between the third lesion region of the mn second segmented images and the first lesion region of the corresponding mn first CTPA images is less than the second difference threshold, wherein the third difference threshold is smaller than the first difference threshold.
[0019] In some possible implementations, the step of the difference between the second lesion region of the n segmented images and the first lesion region of the corresponding n first CTPA images being less than a third difference threshold, and the difference between the third lesion region of the mn second segmented images and the first lesion region of the corresponding mn first CTPA images being less than the second difference threshold, includes:
[0020] The difference between the second lesion region of the n segmented images and the first lesion region of the corresponding n first CTPA images is less than a third difference threshold, and the difference between the third lesion region of the mn second segmented images and the first lesion region of the corresponding mn first CTPA images is less than the third difference threshold, wherein the third difference threshold is smaller than the second difference threshold.
[0021] In some possible implementations, each of the first CTPA images and each of the second CTPA images has the same resolution and both have normalized pixel values;
[0022] The m second CTPA images are derived from different CT scanning devices, and the m first CTPA images are obtained by copying the m second CTPA images.
[0023] In some possible implementations, the first difference threshold is the maximum value of the ratio of the number of allowed failed segmentation images to n, wherein the number of failed segmentation images is the first segmentation image in which the second lesion region does not correspond to the first lesion region in the corresponding first CTPA image.
[0024] Secondly, a pulmonary embolism risk stratification system is proposed, including:
[0025] memory,
[0026] The processor connected to the memory, and
[0027] Program instructions stored in the memory and executable by the processor;
[0028] When the processor executes the program instructions, it implements the method as described in the first aspect.
[0029] Thirdly, a computer-readable storage medium is proposed, which stores program instructions that, when executed by a computer device, implement the method described in the first aspect. Attached Figure Description
[0030] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings of the embodiments will be briefly described below. Obviously, the drawings described below only relate to some embodiments of this application, and are not intended to limit this application.
[0031] Figure 1 This is a flowchart of the preparation method of the biomimetic liposome drug delivery system provided in the embodiments of this application.
[0032] Figure 2 This is a synthesis and characterization diagram of the biomimetic liposome drug delivery system provided in the embodiments of this application. Detailed Implementation
[0033] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the described embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. It is understood that, without conflict, some technical means of the various embodiments described herein can be substituted for or combined with each other.
[0034] In the description of this application and the claims, the terms "first," "second," etc., are used only to distinguish the described objects and have no sequential or technical meaning. Therefore, objects specified with "first," "second," etc., may explicitly or implicitly include one or more of those objects. Furthermore, the words "one" or "a" do not indicate a quantity limitation, but rather indicate the presence of at least one, while "multiple" indicates not less than two.
[0035] Figure 1 This application illustrates a pulmonary embolism risk stratification method based on an embodiment of the present application. The method is based on a fused image segmentation model and a prediction model, and includes:
[0036] S101, training the image segmentation model.
[0037] Image segmentation models can be based on the U-Net architecture and are referred to as U-Net models. U-Net is a classic image segmentation network structure based on Convolutional Neural Networks (CNNs). It achieves pixel-by-pixel classification of images through an encoder-decoder structure and is widely used in medical image analysis due to its excellent image segmentation performance. To verify the performance of the U-Net model in the pulmonary embolism segmentation task, the inventors compared it with other common medical image segmentation models, including FCN, SegNet, DeepLabV3, Mask R-CNN, and PSPNet. Experimental comparisons showed that the U-Net model performed excellently in terms of segmentation accuracy, recall, and inference speed. Therefore, the U-Net model was ultimately selected for the automatic pulmonary embolism segmentation task.
[0038] like Figure 2 As shown, in some embodiments, step S101 may specifically include:
[0039] S201, acquire m first CTPA images and m second CTPA images with the same image content as the m first CTPA images, wherein each first CTPA image is additionally marked with a first lesion region relative to the corresponding second CTPA image, and m is an integer not less than 100.
[0040] In one example, the first example, 300 (one case of m) second CTPA images are obtained from clinical trials or open-source databases, and 300 first CTPA images are obtained by copying these 300 second CTPA images. Specifically, 300 original CTPA images can be obtained from clinical trials or open-source databases. These 300 original CTPA images are then subjected to resolution adjustment, pixel value normalization, and data augmentation to obtain 300 second CTPA images with the same resolution and normalized pixel values. These 300 second CTPA images are then copied using computer equipment to obtain 300 second CTPA copy images. A researcher with experience in imaging diagnosis and image processing then uses 3D Slicer software to delineate the lesion regions (i.e., the first lesion regions) of these 300 second CTPA copy images. The lesion delineation of all second CTPA copy images is reviewed by another clinician to ensure the reliability of the delineation data. Thus, 300 second CTPA images are obtained. It is understandable that the 300 first CTPA images and the 300 second CTPA images have the same image content in a one-to-one correspondence, but each first CTPA image is additionally marked with the lesion area relative to the corresponding second CTPA image.
[0041] Since each first CTPA image and each second CTPA image have the same resolution and normalized pixel values, and have undergone data augmentation, the consistency and generalization ability of the model input data are optimized.
[0042] In some embodiments, the m second CTPA images are obtained from different CT scanning devices. Using CTPA images obtained from scanning with different CT scanning devices to train the image segmentation model can improve the adaptability of the image segmentation model to the CTPA images to be segmented.
[0043] S202, input n of the m second CTPA images into the image segmentation model to obtain n first segmented images with the second lesion region marked by the image segmentation model, where n < m.
[0044] For example, in the first example mentioned above, 240 (80%) images were randomly selected from 300 second CTPA images and input into the image segmentation model to be trained, resulting in 240 first segmented images with the lesion region, i.e. the second lesion region, output by the image segmentation model.
[0045] S203, if the difference between the second lesion region of the n first segmented images and the first lesion region of the corresponding n first CTPA images is greater than the first difference threshold, then adjust the parameters of the image segmentation model until the difference between the second lesion region of the n segmented images and the first lesion region of the corresponding n first CTPA images is less than the first difference threshold.
[0046] The first difference threshold can be the maximum value of the ratio of the number of allowed failed segmentation images to the total number of first segmentation images (i.e., n), where the number of failed segmentation images refers to the first segmentation images whose second lesion region does not correspond to the first lesion region in the corresponding first CTPA image.
[0047] For example, in the first example mentioned above, if it is confirmed that more than 5% (i.e., 12) of the 240 first segmented images output by the image segmentation model do not correspond to the first lesion region of the corresponding first CTPA image, then the image segmentation model is considered to have converged. Therefore, after adjusting the parameters of the image segmentation model, the operation of S202 is repeated until less than 12 of the second lesion regions of the first segmented images do not correspond to the first lesion regions of the corresponding first CTPA images, and the parameters of the image segmentation model are maintained at the last adjusted parameters.
[0048] Since the first lesion region of the first CTPA image in this embodiment is marked before step S202, that is, before the first segmented image is obtained, the difference between the second lesion region of the first segmented image and the second lesion region of the first segmented image is not affected by certain subjective factors when comparing (especially by manual comparison) the difference between the second lesion region of the first segmented image and the second lesion region of the first segmented image. Therefore, the training efficiency of the model is better.
[0049] When adjusting the parameters of an image segmentation model, the model loss function can be optimized based on the Dice coefficient. By introducing the Dice coefficient to optimize the segmentation model, the problems of small pulmonary embolism lesion area and class imbalance can be addressed, thereby improving segmentation accuracy.
[0050] S204, input the mn second CTPA images (excluding the n second CTPA images) into the image segmentation model to obtain mn second segmented images with the third lesion region marked by the image segmentation model.
[0051] That is, after training an image segmentation model with parameters that meet the set requirements using n second CTPA images, the remaining mn second CTPA images, corresponding to the model's validation set, are input into the pre-trained image segmentation model. The quality of the mn second segmented images output by the model is then used to verify whether the image segmentation model meets the application requirements. Therefore, the n second CTPA images correspond to the model's training set, and the mn second CTPA images correspond to the model's validation set. In some embodiments, the ratio of n to m can be 80%.
[0052] S205, if the difference between the third lesion region of the aforementioned mn second segmented images and the first lesion region of the corresponding mn first CTPA images is greater than the second difference threshold, then the parameters of the image segmentation model are adjusted until the difference between the second lesion region of the aforementioned n segmented images and the first lesion region of the corresponding n first CTPA images is less than the third difference threshold, and the difference between the third lesion region of the aforementioned mn second segmented images and the first lesion region of the corresponding mn first CTPA images is less than the second difference threshold, wherein the third difference threshold is smaller than the first difference threshold.
[0053] In this embodiment, when the quality of the mn second segmentation images output by the model after parameter adjustment does not meet the requirements, the model is not only trained again based on the previous n second CTPA images, but also trained using other mn second segmentation images. Furthermore, when training the model again based on the n second CTPA images, the previously allowed first difference threshold is reduced to a third difference threshold, thus expending more training resources and time to improve the model's convergence and enhance its segmentation quality. In other words, if the difference between the third lesion region of the mn second segmentation images and the first lesion region of the corresponding mn first CTPA images is less than the second difference threshold in step S205, it is not necessary to adjust the model parameters again, thereby reducing the resources and time required for model training and improving training efficiency.
[0054] The second difference threshold is usually not less than the first difference threshold, however, in some embodiments, the second difference threshold may be equal to the first difference threshold.
[0055] In some embodiments, the statement in S205 that the difference between the second lesion region of the aforementioned n segmented images and the first lesion region of the corresponding n first CTPA images is less than a third difference threshold, and the difference between the third lesion region of the aforementioned mn second segmented images and the first lesion region of the corresponding mn first CTPA images is less than a second difference threshold, can specifically be: the difference between the second lesion region of the aforementioned n segmented images and the first lesion region of the corresponding n first CTPA images is less than a third difference threshold, and the difference between the third lesion region of the aforementioned mn second segmented images and the first lesion region of the corresponding mn first CTPA images is also less than a third difference threshold, wherein the third difference threshold is smaller than the second difference threshold.
[0056] S102, input the third CTPA image into the image segmentation model to obtain the third segmented image with the fourth lesion region marked by the image segmentation model.
[0057] The third CTPA image can be a CTPA image generated in a clinical setting, rather than an image used to train the image segmentation model. That is, after training the image segmentation model, the model can be used to segment CTPA images such as those generated in a clinical setting.
[0058] S103, input the third segmented image into the prediction model to obtain the pulmonary embolism risk stratification prediction result output by the prediction model, wherein the pulmonary embolism risk stratification prediction result is determined by the prediction model based on the radiomics features of the fourth lesion region, and the radiomics features are extracted by the prediction model from the fourth lesion region.
[0059] In this way, by fusing the U-Net architecture image segmentation model with a prediction model that has image omics feature analysis capabilities, automatic segmentation and risk stratification diagnosis of pulmonary embolism can be achieved.
[0060] Image-based features can include embolus volume, CT value (i.e. embolus enhanced CT value), and clot burden score. By extracting features such as coronary artery embolus volume, CT value, and clot burden score through image-based methods, the hierarchical prediction performance of the prediction model can be enhanced.
[0061] Before step S103, the prediction model can be trained using the K-means clustering algorithm based on three radiomics features: embolus volume, CT value, and clot burden score. The K-means algorithm is a clustering analysis algorithm that calculates data clusters by repeatedly taking the mean of the nearest seed point. It can automatically divide data objects into multiple categories or clusters by mining the inherent structure and patterns in the data, even without prior knowledge or label information. Objects within each cluster have high similarity, while objects in different clusters exhibit significant differences. Furthermore, the random forest algorithm is an ensemble learning method that generates more accurate and stable predictions by constructing multiple decision trees and merging them. This method can determine which feature has the greatest impact on the clustering results by evaluating the importance of each feature. Therefore, to evaluate the importance of different imaging features in the risk stratification prediction model, after the K-means clustering analysis is completed, the random forest algorithm can also be used to analyze the importance of the three imaging features—embolus volume, CT value, and clot burden score—in the pulmonary embolism risk stratification prediction model. The random forest algorithm is used to evaluate the contribution of radiomics features to risk stratification prediction and optimize the reliability of the prediction model. The fusion of radiomics features further refines the dimensions of assessment. Combined with K-means clustering analysis and random forest algorithm, it ensures that patients at different risk levels are classified more accurately.
[0062] This application also proposes a pulmonary embolism risk stratification system, including: a memory, a processor connected to the memory, and program instructions stored in the memory and executable by the processor, wherein the processor executes the program instructions to implement the above-described method.
[0063] In addition, embodiments of this application also propose a computer-readable storage medium storing program instructions that, when executed by a computer device, implement the above-described method.
Claims
1. A method for risk stratification of pulmonary embolism, characterized in that, include: The CTPA image is input into the image segmentation model to obtain a segmented image with the lesion region marked by the image segmentation model. The segmented image is input into the prediction model to obtain the pulmonary embolism risk stratification prediction result output by the prediction model, wherein the pulmonary embolism risk stratification prediction result is determined by the prediction model based on the radiomics features of the lesion region, and the radiomics features are extracted by the prediction model from the lesion region.
2. The method according to claim 1, characterized in that, The radiomics features include embolus volume, CT value, and clot burden score.
3. The method according to claim 1, characterized in that, Before inputting the CTPA image into the image segmentation model, the method further includes training the image segmentation model, specifically including: Obtain m first CTPA images and m second CTPA images with the same image content as the m first CTPA images, wherein each first CTPA image is additionally marked with a first lesion region relative to the corresponding second CTPA image, and m is an integer not less than 100; Input n of the m second CTPA images into the image segmentation model to obtain n first segmented images with the second lesion region marked by the image segmentation model, where n < m; If the difference between the second lesion region of the n first segmented images and the first lesion region of the corresponding n first CTPA images is greater than a first difference threshold, then the parameters of the image segmentation model are adjusted until the difference between the second lesion region of the n segmented images and the first lesion region of the corresponding n first CTPA images is less than the first difference threshold, and thereafter... The mn second CTPA images, excluding the n second CTPA images, are input into the image segmentation model to obtain mn second segmented images with the third lesion region marked by the image segmentation model. If the difference between the third lesion region of the mn second segmented images and the first lesion region of the corresponding mn first CTPA images is greater than a second difference threshold, then the parameters of the image segmentation model are adjusted until the difference between the second lesion region of the n segmented images and the first lesion region of the corresponding n first CTPA images is less than the first difference threshold, and the difference between the third lesion region of the mn second segmented images and the first lesion region of the corresponding mn first CTPA images is less than the second difference threshold, wherein the third difference threshold is smaller than the first difference threshold.
4. The method according to claim 3, characterized in that, The second difference threshold is equal to the first difference threshold, and the image segmentation model is a neural network model based on the U-net architecture.
5. The method according to claim 3, characterized in that, The condition that the difference between the second lesion region of the n segmented images and the first lesion region of the corresponding n first CTPA images is less than a first difference threshold, and the difference between the third lesion region of the mn second segmented images and the first lesion region of the corresponding mn first CTPA images is less than a second difference threshold, includes: The difference between the second lesion region of the n segmented images and the first lesion region of the corresponding n first CTPA images is less than a third difference threshold, and the difference between the third lesion region of the mn second segmented images and the first lesion region of the corresponding mn first CTPA images is less than the second difference threshold, wherein the third difference threshold is smaller than the first difference threshold.
6. The method according to claim 5, characterized in that, The condition that the difference between the second lesion region of the n segmented images and the first lesion region of the corresponding n first CTPA images is less than a third difference threshold, and the difference between the third lesion region of the mn second segmented images and the first lesion region of the corresponding mn first CTPA images is less than the second difference threshold, includes: The difference between the second lesion region of the n segmented images and the first lesion region of the corresponding n first CTPA images is less than a third difference threshold, and the difference between the third lesion region of the mn second segmented images and the first lesion region of the corresponding mn first CTPA images is less than the third difference threshold, wherein the third difference threshold is smaller than the second difference threshold.
7. The method according to claim 3, characterized in that, Each of the first CTPA images and each of the second CTPA images has the same resolution and normalized pixel values; The m second CTPA images are derived from different CT scanning devices, and the m first CTPA images are obtained by copying the m second CTPA images.
8. The method according to claim 3, characterized in that, The first difference threshold is the maximum value of the ratio of the number of allowed failed segmentation images to n, wherein the number of failed segmentation images is the first segmentation image in which the second lesion region does not correspond to the first lesion region in the corresponding first CTPA image.
9. A pulmonary embolism risk stratification system, characterized in that, include: memory, The processor connected to the memory, and Program instructions stored in the memory and executable by the processor; When the processor executes the program instructions, it implements the method as described in any one of claims 1 to 8.
10. A computer-readable storage medium storing program instructions, characterized in that, When the program instructions are executed by a computer device, the method described in any one of claims 1 to 9 is implemented.