A method, apparatus, electronic device, and storage medium for determining vascular lesions.

Abstract

The application provides a blood vessel lesion determination method and device, electronic equipment and storage medium. The determination method comprises: inputting acquired no-label images and labeled images as input images into a blood vessel lesion analysis model in a single-alternating input manner; if the input image is a no-label image, training the blood vessel lesion analysis model according to a reconstruction loss function; if the input image is a labeled image, determining whether the reconstruction loss function, a lesion type loss function and a lesion degree loss function simultaneously satisfy a target condition, and if the target condition is satisfied, obtaining a trained blood vessel lesion analysis model; and determining a lesion type result and a lesion degree result of a blood vessel image according to the trained blood vessel lesion analysis model. The technical solution provided by the application can reduce manual labeling operations, reduce the workload of doctors, and ensure the accuracy of blood vessel lesion determination.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a blood vessel lesion determination method and device, electronic equipment and storage medium. BACKGROUND

[0002] With the continuous improvement of people's living standards, the improvement of diet level leads to the multiple occurrence of basic diseases such as obesity, hypertension and hyperlipidemia, which are important factors causing the multiple occurrence of coronary heart disease; then with the development of science and technology, medical imaging technology is also developing rapidly, so that the shooting of CTA is becoming more and more popular. As an important means of coronary heart disease screening, CTA can find signs of blood vessel lesions in the early stage and give reasonable analysis for doctors to make diagnosis and treatment reference, which is of great significance. However, a large number of CTA screening will make the doctors who are already under heavy work pressure even more burdened, therefore, it is a major development direction to assist doctors in blood vessel lesion analysis through deep learning technology.

[0003] At present, deep learning models for learning complex tasks generally require a large amount of data support, and blood vessel lesion analysis models based on deep learning often need a large number of CTA data sets manually labeled by doctors for training. Since the labeling task has high requirements, it often needs experienced doctors to complete, which makes it difficult to obtain labeled data. Therefore, how to determine the blood vessel lesion can reduce the manual labeling operation while ensuring the accuracy of determining the blood vessel lesion, which has become a problem to be solved. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a blood vessel lesion determination method and device, electronic equipment and storage medium, which can train a blood vessel lesion analysis model in a single alternating input manner by using a large number of unlabeled images and a small number of labeled images as input images, so as to obtain a trained blood vessel lesion analysis model that can determine the lesion type result and lesion degree result of the blood vessels in the blood vessel image. This training method without a large number of labeled images can reduce the manual labeling operation and reduce the workload of doctors, and at the same time, the alternating training of unlabeled images and labeled images and the training method of simultaneously satisfying the target conditions according to the reconstruction loss function, the lesion type loss function and the lesion degree loss function also ensure the accuracy of determining the blood vessel lesion.

[0005] The present application mainly includes the following aspects:

[0006] In a first aspect, the embodiments of the present application provide a blood vessel lesion determination method, which comprises:

[0007] obtaining a blood vessel image;

[0008] input the blood vessel image into a blood vessel lesion analysis model, and output a lesion type result and a lesion degree result;

[0009] The blood vessel lesion analysis model is trained through the following steps:

[0010] (A) obtaining a number of unmarked images and a number of marked images with pre-labeled lesion types and lesion degrees, the number of unmarked images being a first preset threshold, and the number of marked images being a second preset threshold, wherein the first preset threshold is a preset multiple of the second preset threshold;

[0011] (B) inputting the unmarked images and the marked images as input images into the blood vessel lesion analysis model in a single-alternating input manner;

[0012] (C) if the input image is an unmarked image, obtaining a reconstruction loss function, training the blood vessel lesion analysis model according to the reconstruction loss function, and returning to step (B);

[0013] (D) if the input image is a marked image, obtaining a reconstruction loss function, a lesion type loss function, and a lesion degree loss function, determining whether the reconstruction loss function, the lesion type loss function, and the lesion degree loss function simultaneously satisfy a target condition, if not, returning to step (B); if the target condition is satisfied, obtaining the trained blood vessel lesion analysis model; the target condition is that the reconstruction loss function is less than a preset reconstruction threshold, the lesion type loss function is less than a preset classification threshold, and the lesion degree loss function is less than a preset degree threshold.

[0014] Further, the blood vessel lesion analysis model comprises a feature extraction module and an image reconstruction module;

[0015] The reconstruction loss function is obtained through the following steps:

[0016] inputting the input image into the feature extraction module for feature extraction to obtain a feature image corresponding to the feature extraction module;

[0017] inputting the feature image into the image reconstruction module for image reconstruction to obtain a target reconstruction image output by the image reconstruction module;

[0018] comparing the target reconstruction image and the input image to obtain a reconstruction loss function.

[0019] Further, the feature extraction module comprises a plurality of extraction network structures, each extraction network structure comprising N convolutional layers and M pooling layers;

[0020] The step of inputting the input image into the feature extraction module for feature extraction to obtain a feature image corresponding to the feature extraction module comprises:

[0021] The input image is input into a first extraction network structure of the feature extraction module, and the first extraction network structure is taken as a current extraction network structure; after feature information of the input image is extracted through N convolutional layers of the current extraction network structure, the extracted feature information is condensed through M pooling layers of the current extraction network structure to obtain a feature image corresponding to the current extraction network structure;

[0022] The next extraction network structure of the first extraction network structure is updated as a current extraction network structure, and a feature image corresponding to the previous extraction network structure is input into the current extraction network structure as an input image to continue to obtain a feature image corresponding to the current extraction network structure, until the last extraction network structure in the feature extraction module is traversed to obtain a feature image corresponding to each extraction network structure in the feature extraction module.

[0023] Further, the image reconstruction module comprises the same number of reconstruction network structures as the feature extraction module, and each reconstruction network structure comprises N convolutional layers and M deconvolutional layers; the step of inputting the feature image into the image reconstruction module for image reconstruction to obtain a target reconstruction image output by the image reconstruction module comprises:

[0024] The feature image corresponding to each extraction network structure in the feature extraction module is input into a reconstruction network structure corresponding to the reconstruction network structure;

[0025] The first reconstruction network structure of the image reconstruction module is taken as a current reconstruction network structure, and a feature image corresponding to the last extraction network structure of the feature extraction module is extracted through N convolutional layers of the current reconstruction network structure, and then up-sampling is performed through M deconvolutional layers of the current reconstruction network structure to obtain a reconstruction image corresponding to the current reconstruction network structure;

[0026] The next reconstruction network structure of the first reconstruction network structure is updated as a current reconstruction network structure, and a reconstruction image corresponding to the previous reconstruction network structure and a feature image corresponding to a corresponding extraction network structure in the feature extraction module are input into the current reconstruction network structure together to continue to obtain a reconstruction image corresponding to the current reconstruction network structure, until the last reconstruction network structure in the image reconstruction module is traversed to obtain a target reconstruction image output by the last reconstruction network structure in the image reconstruction module.

[0027] Furthermore, the vascular lesion analysis model also includes a lesion analysis module; the lesion analysis module includes a lesion classifier and a lesion severity regressor; if the input image is a labeled image, the lesion type loss function and the lesion severity loss function are obtained through the following steps:

[0028] The feature images corresponding to each extraction network structure in the feature extraction module are input into the lesion analysis module for overlay and fusion to obtain the target feature image;

[0029] The target feature image is input into the lesion classifier for classification to obtain the lesion type result. Based on the lesion type result and the pre-labeled lesion types in the labeled image, the lesion type loss function is obtained.

[0030] The target feature image is input into the lesion severity regressor to predict the lesion severity and obtain the lesion severity result. Based on the lesion severity result and the pre-labeled lesion severity in the labeled image, the lesion severity loss function is obtained.

[0031] Furthermore, the number of unlabeled images that is a first preset threshold and the number of labeled images that are pre-labeled with lesion type and lesion degree that is a second preset threshold are both images that have undergone data processing;

[0032] The data processing is performed through the following steps:

[0033] For each labeled image and each unlabeled image, the labeled image and the unlabeled image are resampled according to a preset spatial resolution to obtain labeled images and unlabeled images with uniform spatial resolution;

[0034] The labeled and unlabeled images with uniform spatial resolution are then subjected to uniform brightness within a preset brightness range to obtain labeled and unlabeled images with uniform brightness.

[0035] The labeled and unlabeled images with uniform brightness are subjected to standard score processing to obtain unlabeled images with a number of data processing steps equal to a first preset threshold and labeled images with a number of pre-labeled lesion types and lesion degrees equal to a second preset threshold.

[0036] Furthermore, the step of inputting the vascular image into the vascular lesion analysis model and outputting lesion type and severity results includes:

[0037] The vascular image is input into the vascular lesion analysis model, and features are extracted by the feature extraction module in the vascular lesion analysis model to obtain multiple feature images of the vascular image;

[0038] Multiple feature images of the blood vessel image are processed by the lesion analysis module in the blood vessel lesion analysis model to obtain the lesion type result output by the lesion classifier of the lesion analysis module and the lesion degree result output by the lesion degree regressor of the lesion analysis module.

[0039] Secondly, embodiments of this application also provide a device for determining vascular lesions, the device comprising:

[0040] The acquisition module is used to acquire vascular images;

[0041] The output module is used to input the vascular image into the vascular lesion analysis model and output the lesion type and lesion severity results;

[0042] A training module is used to train a vascular lesion analysis model; the training module includes an acquisition unit, an input unit, a first training unit, and a second training unit.

[0043] The acquisition unit is used to acquire an unlabeled image of a quantity equal to a first preset threshold and a labeled image of a quantity equal to a second preset threshold, which is pre-labeled with lesion type and lesion degree; wherein the first preset threshold is a preset multiple of the second preset threshold;

[0044] The input unit is used to input the unlabeled image and the labeled image as input images into the vascular lesion analysis model in a single, alternating input manner.

[0045] The first training unit is used to obtain a reconstruction loss function when the input image is an unlabeled image, train the vascular lesion analysis model according to the reconstruction loss function, and return to the input unit;

[0046] The second training unit is used to obtain a reconstruction loss function, a lesion type loss function, and a lesion severity loss function when the input image is a labeled image, and to determine whether the reconstruction loss function, the lesion type loss function, and the lesion severity loss function simultaneously satisfy the target conditions. If they do not simultaneously satisfy the target conditions, the training unit proceeds to the input unit. If they simultaneously satisfy the target conditions, the trained vascular lesion analysis model is obtained. The target conditions are that the reconstruction loss function is less than a preset reconstruction threshold, the lesion type loss function is less than a preset classification threshold, and the lesion severity loss function is less than a preset severity threshold.

[0047] Thirdly, embodiments of this application also provide an electronic device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, the steps of the method for determining vascular lesions as described above are performed.

[0048] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the method for determining vascular lesions as described above.

[0049] This application provides a method, apparatus, electronic device, and storage medium for determining vascular lesions. The determination method includes: acquiring vascular images; inputting the vascular images into a vascular lesion analysis model, and outputting lesion type and lesion severity results; wherein the vascular lesion analysis model is trained through the following steps: (A) acquiring unlabeled images of a number equal to a first preset threshold and labeled images of a number equal to a second preset threshold, pre-labeled with lesion type and lesion severity; wherein the first preset threshold is a preset multiple of the second preset threshold; (B) inputting the unlabeled images and the labeled images as input images into the vascular lesion analysis model in a single-image alternating input manner; (C) if the input image is an unlabeled image, The reconstruction loss function is then obtained. The vascular lesion analysis model is trained based on the reconstruction loss function, and the process returns to step (B); (D) If the input image is a labeled image, the reconstruction loss function, lesion type loss function, and lesion severity loss function are obtained. It is determined whether the reconstruction loss function, lesion type loss function, and lesion severity loss function simultaneously satisfy the target conditions. If they do not simultaneously satisfy the target conditions, the process returns to step (B); if they simultaneously satisfy the target conditions, the trained vascular lesion analysis model is obtained; the target conditions are that the reconstruction loss function is less than a preset reconstruction threshold, the lesion type loss function is less than a preset classification threshold, and the lesion severity loss function is less than a preset severity threshold.

[0050] Thus, the technical solution provided in this application can train the vascular lesion analysis model by using a large number of unlabeled images and a small number of labeled images as input images in a single-image alternating input manner. The trained vascular lesion analysis model can determine the lesion type and degree of vascular lesions in vascular images. This training method, which does not require a large number of labeled images, can reduce the manual annotation operation and reduce the workload of doctors. At the same time, the alternating training with unlabeled and labeled images, as well as the training method that simultaneously satisfies the target conditions based on the reconstruction loss function, lesion type loss function, and lesion degree loss function, also ensures the accuracy of vascular lesion determination.

[0051] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 A flowchart illustrating a method for determining vascular lesions provided in an embodiment of this application is shown;

[0054] Figure 2 A flowchart illustrating another method for determining vascular lesions provided in an embodiment of this application is shown;

[0055] Figure 3 A structural diagram of a vascular lesion analysis model provided in an embodiment of this application is shown;

[0056] Figure 4 A schematic diagram of a training vascular lesion analysis model provided in an embodiment of this application is shown;

[0057] Figure 5 This illustration shows one of the structural diagrams of a device for determining vascular lesions provided in an embodiment of this application;

[0058] Figure 6 This is a second structural diagram of a device for determining vascular lesions provided in an embodiment of this application;

[0059] Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the embodiments 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. It should be understood that the drawings in this application are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0061] Furthermore, the described embodiments are merely some, not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0062] To enable those skilled in the art to use the content of this application in conjunction with the specific application scenario of "determination of vascular lesions", the following implementation methods are provided. For those skilled in the art, the general principles defined herein can be applied to other embodiments and application scenarios without departing from the spirit and scope of this application.

[0063] The methods, apparatus, electronic devices, or computer-readable storage media described in this application can be applied to any scenario where it is necessary to determine vascular lesions. This application does not limit the specific application scenario. Any scheme that uses a method, apparatus, electronic device, or storage medium for determining vascular lesions provided in this application is within the protection scope of this application.

[0064] It is worth noting that with the continuous improvement of people's living standards and dietary habits, the incidence of underlying diseases such as obesity, hypertension, and hyperlipidemia has increased, which are important factors contributing to the high incidence of coronary heart disease. Meanwhile, with the rapid development of science and technology, medical imaging technology is also advancing by leaps and bounds, leading to the increasing prevalence of CTA (Coronary Acupuncture and Traumatology). CTA, as an important means of screening for coronary heart disease, can detect signs of vascular lesions at an early stage and provide reasonable analysis for doctors to use as a reference for diagnosis and treatment, which is of great significance. However, the large-scale CTA screening will further increase the burden on doctors who already have heavy workloads. Therefore, using deep learning technology to assist doctors in analyzing vascular lesions is a major direction for development.

[0065] Currently, deep learning models for complex tasks generally require substantial data support. However, training deep learning-based vascular lesion analysis models often necessitates training on large CTA datasets manually annotated by physicians. Due to the high demands of annotation tasks, this often requires experienced doctors, making the acquisition of labeled data difficult. Therefore, how to identify vascular lesions while minimizing manual annotation and ensuring accuracy has become a pressing issue.

[0066] Based on this, this application proposes a method, apparatus, electronic device, and storage medium for determining vascular lesions. The determination method includes: acquiring vascular images; inputting the vascular images into a vascular lesion analysis model, and outputting lesion type and lesion severity results; wherein the vascular lesion analysis model is trained through the following steps: (A) acquiring unlabeled images of a number equal to a first preset threshold and labeled images of a number equal to a second preset threshold, pre-labeled with lesion type and lesion severity; wherein the first preset threshold is a preset multiple of the second preset threshold; (B) inputting the unlabeled images and the labeled images as input images into the vascular lesion analysis model in a single-image alternating input manner; (C) if the input image is an unlabeled image If the input image is a labeled image, then the reconstruction loss function, lesion type loss function, and lesion severity loss function are obtained. It is determined whether the reconstruction loss function, lesion type loss function, and lesion severity loss function simultaneously satisfy the target conditions. If they do not simultaneously satisfy the target conditions, then the process returns to step (B). If they simultaneously satisfy the target conditions, then the trained vascular lesion analysis model is obtained. The target conditions are that the reconstruction loss function is less than a preset reconstruction threshold, the lesion type loss function is less than a preset classification threshold, and the lesion severity loss function is less than a preset severity threshold.

[0067] Thus, the technical solution provided in this application can train the vascular lesion analysis model by using a large number of unlabeled images and a small number of labeled images as input images in a single-image alternating input manner. The trained vascular lesion analysis model can determine the lesion type and degree of vascular lesions in vascular images. This training method, which does not require a large number of labeled images, can reduce the manual annotation operation and reduce the workload of doctors. At the same time, the alternating training with unlabeled and labeled images, as well as the training method that simultaneously satisfies the target conditions based on the reconstruction loss function, lesion type loss function, and lesion degree loss function, also ensures the accuracy of vascular lesion determination.

[0068] To facilitate understanding of this application, the technical solutions provided in this application will be described in detail below with reference to specific embodiments.

[0069] Please see Figure 1 , Figure 1 A flowchart illustrating a method for determining vascular lesions provided in an embodiment of this application is shown below. Figure 1 As shown, the determination method includes:

[0070] S101. Acquire vascular images;

[0071] In this step, the vascular images can be CTA images of the blood vessels to help detect signs of vascular lesions at an early stage.

[0072] S102. Input the vascular image into the vascular lesion analysis model and output the lesion type and lesion severity results;

[0073] It should be noted that the vascular lesion analysis model needs to be trained before performing step S102. The vascular image is input into the trained vascular lesion analysis model, which outputs the lesion type and severity results. Please refer to [link to relevant documentation]. Figure 2 , Figure 2 A flowchart illustrating another method for determining business metrics provided in an embodiment of this application is shown below. Figure 2 As shown, the vascular lesion analysis model is trained through the following steps:

[0074] S201. Acquire unlabeled images with a quantity equal to a first preset threshold and labeled images with a quantity equal to a second preset threshold, pre-labeled with lesion type and lesion degree.

[0075] In this step, the first preset threshold is a preset multiple of the second preset threshold, and the preset multiple is pre-set based on historical experience or experimental data. Before training the vascular lesion analysis model, a small dataset is typically manually labeled to indicate the type and severity of vascular lesions, resulting in labeled images that serve as the gold standard for the deep learning network model to learn from.

[0076] Here, model training is based on semi-supervised learning (SSL), a key research area in pattern recognition and machine learning, combining supervised and unsupervised learning. SSL uses a large amount of unlabeled data (e.g., unlabeled images) along with labeled data (e.g., labeled images) for pattern recognition. Using SSL requires minimal personnel while achieving high accuracy, hence its increasing popularity. Deep learning models for complex tasks generally require substantial data support, often necessitating experienced experts in the field to manually create datasets. This is particularly true in vascular lesion analysis, where the scarcity of medical resources and the need for extensive data labeling often present a contradiction. Therefore, a high-quality, large-scale training dataset is extremely valuable and rare. Conversely, with the advancement of science and technology, CTA examinations are becoming increasingly widespread, with tens or even hundreds of thousands of CTA scans performed daily in my country. This means the raw data volume is more than sufficient; however, this data is unlabeled and difficult to utilize for supervised learning. Therefore, by making full use of a large amount of unlabeled data and combining it with some labeled data, semi-supervised training of deep learning network models can be completed. This allows the training of the model to be supported by rich information from large datasets, while also being constrained by key information (labeled data), making the model more robust in its progress toward solving the target task.

[0077] It should be noted that both the number of unlabeled images (the first preset threshold) and the number of labeled images (the second preset threshold) with pre-labeled lesion types and degrees are images that have undergone data processing.

[0078] The data processing is performed through the following steps:

[0079] S2011. For each labeled image and each unlabeled image, the labeled image and the unlabeled image are resampled according to a preset spatial resolution to obtain labeled images and unlabeled images with uniform spatial resolution.

[0080] S2012. The marked image and the unmarked image with uniform spatial resolution are uniformly brightened according to a preset brightness range to obtain marked image and unmarked image with uniform brightness.

[0081] S2013. Perform standard score processing on the labeled and unlabeled images with uniform image brightness to obtain unlabeled images with a number of data processing steps equal to a first preset threshold and labeled images with a number of pre-labeled lesion types and lesion degrees equal to a second preset threshold.

[0082] This step involves organizing and fusing the sample images obtained from the mobile phone. This mainly involves combining labeled data (tagged images) and unlabeled data (unlabeled images) into a single dataset after unified data processing.

[0083] For example, approximately 20,000 vascular image data (sample images) were collected. This included some semi-automatic processing of CTA by doctors before issuing CTA reports, mainly including the extraction of the vessel centerline and qualitative analysis of patients. For instance, in the 20,000 data points, the first and second preset thresholds were pre-set based on experimental data or historical experience. If the first preset threshold was set to 19,000 and the second preset threshold was set to 1,000, then approximately 1,000 data points were randomly selected from the 20,000 data points. Based on the vessel centerline confirmed by the doctor and the issued CTA report, each lesion in the vessel was manually located and analyzed. The location of the lesion on the vessel centerline, as well as the lesion type and degree at each location, were marked. The approximately 1,000 marked data points were used as labeled images. The remaining 19,000 data points were not manually marked and were used as unlabeled images.

[0084] As an example, considering that these examples come from different hospitals and were taken by CT scans from different manufacturers, there are some objective biases in the default parameters. To avoid the influence of these biases, we will perform some preprocessing on this data before training, including standardizing the spatial resolution, such as resampling all CT scans to 0.5×0.5×1mm. 3 (Preset spatial resolution) To preserve the original CT values ​​rather than adding new CT values ​​for learning the original features, nearest neighbor interpolation was used instead of other linear interpolation methods. To avoid problems caused by inconsistent minimum values ​​in some CTA data (e.g., some have a minimum of -3000, while others have a minimum of -1000), CT values ​​exceeding a certain range were set to fixed values. For example, values ​​<-1000 were uniformly set to -1000, and values ​​>1000 were uniformly set to 1000, thus obtaining labeled and unlabeled images with uniform brightness. Finally, to eliminate some data distribution bias, Z-score processing (standard score processing) was used to process the data into a normal distribution (subtracting the mean and dividing by the variance), resulting in 19,000 unlabeled images (first preset threshold) and 1,000 labeled images with pre-labeled lesion types and degrees (second preset threshold).

[0085] S202. Input the unlabeled image and the labeled image as input images into the vascular lesion analysis model in a single, alternating input manner;

[0086] In this step, the vascular lesion analysis model is trained by inputting labeled and unlabeled images into the vascular lesion analysis model in an alternating manner.

[0087] For example, the total dataset is 20,000 images, including 1,000 labeled images and 19,000 unlabeled images. To ensure fairness in training, a training method alternating between labeled and unlabeled images is adopted. Specifically, in the first training iteration, one image is randomly selected from the 1,000 labeled images to enter labeled training mode. In the second training iteration, one image is randomly selected from the 19,000 unlabeled images to enter unlabeled training mode. In the third iteration, the dataset returns to the 1,000 labeled images to enter labeled training mode, and this alternating training process is repeated to obtain a trained vascular lesion analysis model.

[0088] S203. If the input image is an unlabeled image, then the reconstruction loss function is obtained, the vascular lesion analysis model is trained according to the reconstruction loss function, and the process returns to step S202.

[0089] In this step, the vascular lesion analysis model includes a feature extraction module, an image reconstruction module, and a lesion analysis module. If the input image is unlabeled, it enters the unlabeled training mode, which only uses the feature extraction module and the image reconstruction module. Here, the reconstruction loss function is obtained through the following steps:

[0090] S2031. Input the input image into the feature extraction module for feature extraction to obtain the feature image corresponding to the feature extraction module;

[0091] In this step, the feature extraction module includes multiple extraction network structures, each of which includes N convolutional layers and M pooling layers; the step of inputting the input image into the feature extraction module for feature extraction to obtain the feature image corresponding to the feature extraction module includes:

[0092] (1) Input the input image into the first extraction network structure of the feature extraction module, and use the first extraction network structure as the current extraction network structure. After the N convolutional layers of the current extraction network structure extract the feature information of the input image, the M pooling layers of the current extraction network structure are used to condense the extracted feature information to obtain the feature image corresponding to the current extraction network structure.

[0093] (2) Update the next extraction network structure of the first extraction network structure to the current extraction network structure, and input the feature image corresponding to the previous extraction network structure as the input image into the current extraction network structure, and continue to obtain the feature image corresponding to the current extraction network structure until the last extraction network structure in the feature extraction module is traversed, so as to obtain the feature image corresponding to each extraction network structure in the feature extraction module.

[0094] For example, please refer to Figure 3 , Figure 3 A structural diagram of a vascular lesion analysis model provided in an embodiment of this application is shown below. Figure 3 As shown, the vascular lesion analysis model consists of three parts: a feature extraction module, an image reconstruction module, and a lesion analysis module. The feature extraction module consists of three extraction network structures, each containing two convolutional layers (N=2) and one pooling layer (M=1). It is important to note that the pooling layer downsamples the x and y axes of the input image by 1 / 2, while keeping the z-axis unchanged. The feature dimension in the z-direction of the extraction network structure always remains consistent with the z-axis of the input image to identify the lesion status at each point on the vessel centerline along the z-axis. In each extraction network structure, the two convolutional layers are responsible for extracting basic features. The three network blocks from top to bottom have different functions for their convolutional layers because they extract different types of features. The convolutional layers in extraction network structure 1 are biased towards local information such as texture and edges, while the convolutional layers in extraction network structure 2 tend to extract higher-level information such as shape and grayscale. The convolutional layers in extraction network structure 3 are more abstract, involving high-dimensional abstract information of lesion features. The pooling layers in each extraction network structure are responsible for condensing the feature information extracted by the convolutional layers and passing it to the next extraction network structure.

[0095] S2032. Input the feature image into the image reconstruction module to perform image reconstruction, and obtain the target reconstructed image output by the image reconstruction module;

[0096] In this step, the image reconstruction module includes the same number of reconstruction network structures as the feature extraction module, with each reconstruction network structure comprising N convolutional layers and M deconvolutional layers; the step of inputting the feature image into the image reconstruction module for image reconstruction to obtain the target reconstructed image output by the image reconstruction module includes:

[0097] (1) Input the feature image corresponding to each extracted network structure in the feature extraction module into the reconstruction network structure corresponding to the reconstruction network structure;

[0098] (2) The first reconstruction network structure of the image reconstruction module is taken as the current reconstruction network structure, and the feature image corresponding to the last extraction network structure of the feature extraction module is subjected to N convolutional layers of the current reconstruction network structure for feature extraction, and then subjected to M deconvolutional layers of the current reconstruction network structure for upsampling to obtain the reconstruction image corresponding to the current reconstruction network structure.

[0099] (3) Update the next reconstruction network structure of the first reconstruction network structure to the current reconstruction network structure, and input the reconstruction image corresponding to the previous reconstruction network structure and the feature image corresponding to the extraction network structure in the feature extraction module into the current reconstruction network structure to continue to obtain the reconstruction image corresponding to the current reconstruction network structure until the last reconstruction network structure in the image reconstruction module is traversed to obtain the target reconstruction image output by the last reconstruction network structure of the image reconstruction module.

[0100] In this step, such as Figure 3 As shown, the structure of the image reconstruction module is consistent with that of the feature extraction module, consisting of multiple reconstruction network structures. The number of reconstruction network structures is equal to the number of extraction network structures. Each reconstruction network structure has N convolutional layers and M deconvolutional layers. In each reconstruction network structure, the input of the first convolutional layer is the output image of the previous reconstruction network structure and the feature image transmitted from the corresponding extraction network structure of the feature extraction module. The purpose is to preserve the basic features to the greatest extent during reconstruction. The role of the deconvolutional layer is to upsample the features fused by the convolutional layers, so that the size of the output reconstructed image is continuously restored to the size of the original image (the input image of the first extraction network structure) without losing details.

[0101] S2033. Compare the target reconstructed image with the input image to obtain the reconstruction loss function.

[0102] In this step, the reconstruction loss function can be the mean squared error (MSE). The target reconstructed image and the input image are compared to obtain the similarity. The value of the reconstruction loss function is determined based on the similarity. The greater the similarity, the closer the value of the reconstruction loss function is to 0.

[0103] Here, in step S203, a preset reconstruction threshold can be set in advance based on experimental data or historical experience. When the value of the reconstruction loss function is not less than the preset reconstruction threshold, the weight parameters of the vascular lesion analysis model are modified to train the vascular lesion analysis model. Then, the process returns to step S202, and the labeled image is re-inputted to enter the labeled training mode.

[0104] S204. If the input image is a labeled image, a reconstruction loss function, a lesion type loss function, and a lesion severity loss function are obtained. It is determined whether the reconstruction loss function, the lesion type loss function, and the lesion severity loss function simultaneously satisfy the target conditions. If they do not simultaneously satisfy the target conditions, the process returns to step S202. If they simultaneously satisfy the target conditions, the trained vascular lesion analysis model is obtained. The target conditions are that the reconstruction loss function is less than a preset reconstruction threshold, the lesion type loss function is less than a preset classification threshold, and the lesion severity loss function is less than a preset severity threshold.

[0105] It should be noted that the vascular lesion analysis model also includes a lesion analysis module; the lesion analysis module includes a lesion classifier and a lesion severity regressor; if the input image is a labeled image, the lesion type loss function and the lesion severity loss function are obtained through the following steps:

[0106] S2041. Input the feature image corresponding to each extraction network structure in the feature extraction module into the lesion analysis module for superposition and fusion to obtain the target feature image;

[0107] S2042. Input the target feature image into the lesion classifier for classification to obtain the lesion type result. Based on the lesion type result and the pre-labeled lesion types in the labeled image, obtain the lesion type loss function.

[0108] S2043. Input the target feature image into the lesion degree regressor to predict the lesion degree and obtain the lesion degree result. Based on the lesion degree result and the lesion degree pre-labeled in the labeled image, obtain the lesion degree loss function.

[0109] In this step, the lesion analysis module mainly fuses basic feature information from different scales from the feature extraction module, and then analyzes the lesions, involving lesion classification and lesion severity prediction. The network structure of the lesion analysis module mainly consists of two parallel fully connected layers, which serve as the lesion classifier and the lesion severity regressor, respectively.

[0110] For example, the feature extraction module has three extraction network structures: extraction network structure 1 outputs feature image 1, extraction network structure 2 outputs feature image 2, and extraction network structure 3 outputs feature image 3. Feature images 1, 2, and 3 are input into the lesion analysis module for overlay and fusion to obtain the target feature image. The target feature image is then subjected to lesion analysis and input into the lesion classifier and lesion severity regressor, respectively. The lesion classifier predicts the lesion type corresponding to each point of the vessel centerline in the input labeled image, and the lesion severity regressor predicts the lesion severity corresponding to each point of the vessel centerline in the input labeled image. Based on the predicted lesion type and the corresponding pre-labeled lesion type in the input labeled image, a lesion type loss function is obtained to quantify the difference between the predicted lesion type distribution and the actual lesion type distribution. Based on the predicted lesion severity and the corresponding pre-labeled lesion severity in the input labeled image, a lesion severity loss function is obtained to quantify the difference between the predicted lesion severity distribution and the actual lesion severity distribution.

[0111] Specifically, when the input image is a labeled image, the labeled training mode is entered. Two loss functions constrain the network parameters: the reconstruction loss function (Loss) and the target reconstructed image (MSE) compared with the input labeled image. The steps for obtaining the target reconstructed image and the reconstruction loss function are the same as those in the unlabeled training mode, and will not be repeated here. The lesion analysis loss function (Loss) has two parts: the lesion type loss function (constrained by cross-entropy) and the lesion type loss function (constrained by regression loss function L1Loss). Pre-set classification thresholds and severity thresholds are set based on experimental data or historical experience. A well-trained vascular lesion analysis model is obtained only when the reconstruction loss function is less than the preset reconstruction threshold, the lesion type loss function is less than the preset classification threshold, and the lesion severity loss function is less than the preset severity threshold. Otherwise, the weight parameters of the vascular lesion analysis model are adjusted to train the model, and the process returns to step S202, where the unlabeled image is input into the vascular lesion analysis model to enter the unlabeled training mode.

[0112] It should be noted that step S102, which involves inputting the vascular image into the vascular lesion analysis model and outputting the lesion type and severity results, includes:

[0113] S1021. Input the blood vessel image into the blood vessel lesion analysis model, and perform feature extraction through the feature extraction module in the blood vessel lesion analysis model to obtain multiple feature images of the blood vessel image;

[0114] S1022. Multiple feature images of the blood vessel image are processed by the lesion analysis module in the blood vessel lesion analysis model to perform lesion analysis, and the lesion type result output by the lesion classifier of the lesion analysis module and the lesion degree result output by the lesion degree regressor of the lesion analysis module are obtained.

[0115] In this step, the trained vascular lesion analysis model is applied to the scenario of identifying vascular lesions. In practical application, it is no longer necessary to reconstruct the input vascular image. Therefore, the image reconstruction module can be "removed" from the framework of the vascular lesion analysis model, making the model more lightweight. So, in the application phase, after the vascular image is input into the vascular lesion analysis model, multiple feature images are obtained through the feature extraction module. These multiple feature images are then fused using the multi-scale feature fusion of the lesion analysis module to obtain the target feature image. The final lesion analysis can then be performed on the target feature image, with the lesion classifier outputting the lesion type result and the lesion severity regressor outputting the lesion severity result.

[0116] For example, please refer to Figure 3 , Figure 3 This is a structural diagram of a vascular lesion analysis model provided in an embodiment of this application, such as... Figure 3As shown, during the training of the vascular lesion analysis model, unlabeled and labeled images are used as input images. For each input image, the input image is fed into the extraction network structure 1 of the feature extraction module. After passing through the convolutional and pooling layers of the extraction network structure 1, feature image 1 is obtained. Feature image 1 is then fed into the reconstruction network structure 3 of the image reconstruction module and the extraction network structure 2 of the feature extraction module. After passing through the convolutional and pooling layers of the extraction network structure 2, feature image 2 is obtained. Feature image 2 is then fed into the extraction network structure 3 of the feature extraction module within the reconstruction network structure 2 of the image reconstruction module. After passing through the convolutional and pooling layers of the extraction network structure 3, feature image 3 is obtained. Feature image 3 is then fed into the reconstruction network structure 1 of the image reconstruction module. After passing through the convolutional and deconvolutional layers of the reconstruction network structure 1, reconstructed image 1 is obtained. Reconstructed image 1 and feature image 2 are then fed through the convolutional and deconvolutional layers of the reconstruction network structure 2 to obtain reconstructed image 2. Reconstructed image 2 and feature image 1 are then fed through the convolutional and deconvolutional layers of the reconstruction network structure 3 to obtain the target reconstructed image. Here, the process of obtaining the target reconstructed image is the same whether the input image is an unlabeled image or a labeled image. If the input image is a labeled image, the obtained feature image 1, feature image 2 and feature image 3 need to be input into the lesion analysis module. After multi-scale feature fusion, feature image 1, feature image 2 and feature image 3 are superimposed to obtain the target feature image. The target feature image is simultaneously input into the lesion classifier and the lesion degree regressor for lesion analysis, and the lesion type result output by the lesion classifier and the lesion degree result output by the lesion degree regressor are obtained.

[0117] For example, please refer to Figure 4 , Figure 4 This is a schematic diagram of a training vascular lesion analysis model provided in an embodiment of this application, as shown below. Figure 4As shown, the vascular lesion analysis model is trained by alternating between labeled and unlabeled training modes. In labeled training mode, labeled images are input to the feature extraction module of the vascular lesion analysis model for feature extraction. Multiple feature images are then input to the image reconstruction module and the lesion analysis module, respectively. The image reconstruction module reconstructs the image based on the received feature images to obtain the target reconstructed image and derives a reconstruction loss function. The lesion analysis module performs lesion analysis based on the received feature images, obtaining lesion type loss functions and lesion severity loss functions. It then determines whether the following conditions are met simultaneously: the reconstruction loss function is less than a preset reconstruction threshold, the lesion type loss function is less than a preset classification threshold, and the lesion severity loss function is less than a preset severity threshold. If all three conditions are met, the training process ends, and the trained vascular lesion analysis model is obtained. Otherwise, the weight parameters of the vascular lesion analysis model are adjusted, and the model enters the unlabeled training mode. In unlabeled training mode, unlabeled images are input... Feature extraction is performed in the feature extraction module of the vascular lesion analysis model. The resulting multiple feature images are then input into the image reconstruction module. The image reconstruction module performs image reconstruction based on the received feature images to obtain the target reconstructed image. A reconstruction loss function is obtained based on the target reconstructed image. It is then determined whether the reconstruction loss function is less than a preset reconstruction threshold. If so, it is checked whether the lesion type loss function and lesion severity loss function obtained in the previous labeled training mode simultaneously satisfy the conditions that the lesion type loss function is less than a preset classification threshold and the lesion severity loss function is less than a preset severity threshold. If both conditions are met, the training process ends, and the trained vascular lesion analysis model is obtained. If they are not simultaneously met, the model enters the labeled training mode. If the reconstruction loss function is not less than the preset reconstruction threshold, the weight parameters of the vascular lesion analysis model are adjusted, and the model enters the labeled training mode again. This alternating training is repeated until the conditions that the reconstruction loss function, lesion type loss function, and lesion severity loss function are all simultaneously satisfied are met.

[0118] Here, by fully utilizing a large amount of unlabeled data and combining it with some labeled data, semi-supervised learning is performed on the model. This leverages the information advantages of big data and the task-oriented approach of the labeled data, providing strong information support for solving the target task. Simultaneously, the model employs a multi-pathway network structure for training, cleverly mixing labeled and unlabeled data for synchronous training. Furthermore, the model utilizes multi-scale feature fusion technology, integrating rich features from different scales across multiple pathways into the final target task (vascular lesion analysis). By alternately training the vascular lesion analysis model with a large amount of unlabeled data (unlabeled images) and a small amount of labeled data (labeled images), this method fully utilizes the rich feature information of unlabeled data to assist the limited amount of labeled data in better completing the target task (vascular lesion analysis). Based on the extensive use of unlabeled raw CTA, the deep learning network model is trained alternately, not limited by the operating system, allowing for fully automated and rapid computation. This fully utilizes the multi-pathway network model, achieving high accuracy, strong interpretability, and good robustness. Experiments demonstrate that the training results of this vascular lesion analysis model are superior to those trained using only labeled data.

[0119] This application provides a method for determining vascular lesions, the method comprising: acquiring vascular images; inputting the vascular images into a vascular lesion analysis model, and outputting lesion type and lesion severity results; wherein the vascular lesion analysis model is trained through the following steps: (A) acquiring unlabeled images of a number equal to a first preset threshold and labeled images of a number equal to a second preset threshold, wherein the first preset threshold is a preset multiple of the second preset threshold; (B) inputting the unlabeled images and the labeled images as input images into the vascular lesion analysis model in a single-image alternating input manner; (C) if the input image is an unlabeled image, then obtaining a reconstructed lesion... (B) If the input image is a labeled image, the reconstruction loss function, lesion type loss function, and lesion severity loss function are obtained. It is determined whether the reconstruction loss function, lesion type loss function, and lesion severity loss function simultaneously satisfy the target conditions. If they do not simultaneously satisfy the target conditions, the process returns to step (B). If they simultaneously satisfy the target conditions, the trained vascular lesion analysis model is obtained. The target conditions are that the reconstruction loss function is less than a preset reconstruction threshold, the lesion type loss function is less than a preset classification threshold, and the lesion severity loss function is less than a preset severity threshold.

[0120] Thus, the technical solution provided in this application can train the vascular lesion analysis model by using a large number of unlabeled images and a small number of labeled images as input images in a single-image alternating input manner. The trained vascular lesion analysis model can determine the lesion type and degree of vascular lesions in vascular images. This training method, which does not require a large number of labeled images, can reduce the manual annotation operation and reduce the workload of doctors. At the same time, the alternating training with unlabeled and labeled images, as well as the training method that simultaneously satisfies the target conditions based on the reconstruction loss function, lesion type loss function, and lesion degree loss function, also ensures the accuracy of vascular lesion determination.

[0121] Based on the same concept, this application also provides a device for determining vascular lesions, corresponding to the method for determining vascular lesions provided in the above embodiments. Since the principle of the device in this application is similar to the method for determining vascular lesions in the above embodiments, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.

[0122] Please see Figure 5 , Figure 6 , Figure 5 This is one of the structural diagrams of a device for determining vascular lesions provided in an embodiment of this application. Figure 6 This is a second structural diagram of a device for determining vascular lesions provided in an embodiment of this application. Figure 5 As shown, the determining device 510 includes:

[0123] Acquisition module 511 is used to acquire blood vessel images;

[0124] Output module 512 is used to input the vascular image into the vascular lesion analysis model and output the lesion type result and lesion degree result;

[0125] Training module 513 is used to train the vascular lesion analysis model; the training module includes acquisition unit 5131, input unit 5132, first training unit 5133 and second training unit 5134.

[0126] The acquisition unit 5131 is used to acquire unlabeled images of a quantity equal to a first preset threshold and labeled images of a quantity equal to a second preset threshold, which are pre-labeled with lesion type and lesion degree; wherein, the first preset threshold is a preset multiple of the second preset threshold;

[0127] The input unit 5132 is used to input the unlabeled image and the labeled image as input images into the vascular lesion analysis model in a single, alternating input manner.

[0128] The first training unit 5133 is used to obtain a reconstruction loss function when the input image is an unlabeled image, train the vascular lesion analysis model according to the reconstruction loss function, and return to the input unit;

[0129] The second training unit 5134 is used to obtain a reconstruction loss function, a lesion type loss function, and a lesion severity loss function when the input image is a labeled image, and to determine whether the reconstruction loss function, the lesion type loss function, and the lesion severity loss function simultaneously satisfy a target condition. If they do not simultaneously satisfy the target condition, the training unit is entered; if they simultaneously satisfy the target condition, the trained vascular lesion analysis model is obtained. The target condition is that the reconstruction loss function is less than a preset reconstruction threshold, the lesion type loss function is less than a preset classification threshold, and the lesion severity loss function is less than a preset severity threshold.

[0130] Optionally, the vascular lesion analysis model includes a feature extraction module and an image reconstruction module; when the first training unit 5133 and the second training unit 5134 are used to obtain the reconstruction loss function, the first training unit 5133 and the second training unit 5134 are specifically used for:

[0131] The input image is input to the feature extraction module for feature extraction to obtain the feature image corresponding to the feature extraction module;

[0132] The feature image is input into the image reconstruction module for image reconstruction, and the target reconstructed image output by the image reconstruction module is obtained.

[0133] The target reconstructed image is compared with the input image to obtain the reconstruction loss function.

[0134] Optionally, the feature extraction module includes multiple extraction network structures, each including N convolutional layers and M pooling layers; the vascular lesion analysis model includes a feature extraction module and an image reconstruction module; when the first training unit 5133 and the second training unit 5134 are used to input the input image into the feature extraction module for feature extraction to obtain the feature image corresponding to the feature extraction module, the first training unit 5133 and the second training unit 5134 are specifically used for:

[0135] The input image is input into the first extraction network structure of the feature extraction module, and the first extraction network structure is used as the current extraction network structure. After the N convolutional layers of the current extraction network structure extract the feature information of the input image, the extracted feature information is then condensed through the M pooling layers of the current extraction network structure to obtain the feature image corresponding to the current extraction network structure.

[0136] The next extraction network structure of the first extraction network structure is updated to the current extraction network structure, and the feature image corresponding to the previous extraction network structure is used as the input image to the current extraction network structure. The feature image corresponding to the current extraction network structure is obtained again until the last extraction network structure in the feature extraction module is traversed, and the feature image corresponding to each extraction network structure in the feature extraction module is obtained.

[0137] Optionally, the image reconstruction module includes the same number of reconstruction network structures as the feature extraction module, each reconstruction network structure including N convolutional layers and M deconvolutional layers; when the first training unit 5133 and the second training unit 5134 are used to input the feature image into the image reconstruction module for image reconstruction to obtain the target reconstructed image output by the image reconstruction module, the first training unit 5133 and the second training unit 5134 are specifically used for:

[0138] The feature image corresponding to each extracted network structure in the feature extraction module is input into the reconstruction network structure corresponding to the reconstruction network structure;

[0139] The first reconstruction network structure of the image reconstruction module is taken as the current reconstruction network structure. The feature image corresponding to the last extraction network structure of the feature extraction module is passed through N convolutional layers of the current reconstruction network structure to extract features, and then passed through M deconvolutional layers of the current reconstruction network structure for upsampling to obtain the reconstructed image corresponding to the current reconstruction network structure.

[0140] The next reconstruction network structure of the first reconstruction network structure is updated to the current reconstruction network structure. The reconstructed image corresponding to the previous reconstruction network structure and the feature image corresponding to the extraction network structure in the feature extraction module are input into the current reconstruction network structure. The reconstruction image corresponding to the current reconstruction network structure is obtained again until the last reconstruction network structure in the image reconstruction module is traversed, and the target reconstruction image output by the last reconstruction network structure of the image reconstruction module is obtained.

[0141] Optionally, the vascular lesion analysis model further includes a lesion analysis module; the lesion analysis module includes a lesion classifier and a lesion severity regressor; when the second training unit 5134 is used to obtain the lesion type loss function and the lesion severity loss function, the second training unit 5134 is specifically used for:

[0142] The feature images corresponding to each extraction network structure in the feature extraction module are input into the lesion analysis module for overlay and fusion to obtain the target feature image;

[0143] The target feature image is input into the lesion classifier for classification to obtain the lesion type result. Based on the lesion type result and the pre-labeled lesion types in the labeled image, the lesion type loss function is obtained.

[0144] The target feature image is input into the lesion severity regressor to predict the lesion severity and obtain the lesion severity result. Based on the lesion severity result and the pre-labeled lesion severity in the labeled image, the lesion severity loss function is obtained.

[0145] Optional, such as Figure 6 As shown, the number of unlabeled images (a first preset threshold) and the number of labeled images (a second preset threshold) pre-labeled with lesion type and severity are both images that have undergone data processing; the training module 513 also includes a processing unit 5135, which is used for:

[0146] For each labeled image and each unlabeled image, the labeled image and the unlabeled image are resampled according to a preset spatial resolution to obtain labeled images and unlabeled images with uniform spatial resolution;

[0147] The labeled and unlabeled images with uniform spatial resolution are then subjected to uniform brightness within a preset brightness range to obtain labeled and unlabeled images with uniform brightness.

[0148] The labeled and unlabeled images with uniform brightness are subjected to standard score processing to obtain unlabeled images with a number of data processing steps equal to a first preset threshold and labeled images with a number of pre-labeled lesion types and lesion degrees equal to a second preset threshold.

[0149] Optionally, when the output module 512 is used to input the vascular image into the vascular lesion analysis model and output the lesion type result and lesion severity result, the output module 512 is specifically used for:

[0150] The vascular image is input into the vascular lesion analysis model, and features are extracted by the feature extraction module in the vascular lesion analysis model to obtain multiple feature images of the vascular image;

[0151] Multiple feature images of the blood vessel image are processed by the lesion analysis module in the blood vessel lesion analysis model to obtain the lesion type result output by the lesion type analyzer of the lesion analysis module and the lesion degree result output by the lesion degree regressor of the lesion analysis module.

[0152] This application provides a device for determining vascular lesions, comprising: an acquisition module for acquiring vascular images; an output module for inputting the vascular images into a vascular lesion analysis model and outputting lesion type and severity results; and a training module for training the vascular lesion analysis model. The training module includes an acquisition unit, an input unit, a first training unit, and a second training unit. The acquisition unit acquires unlabeled images of a first preset threshold quantity and labeled images pre-labeled with lesion type and severity of a second preset threshold quantity, wherein the first preset threshold is a preset multiple of the second preset threshold. The input unit inputs the unlabeled images and the labeled images as input images into the vascular lesion analysis model in a single-image alternating input manner. A training unit is used to obtain a reconstruction loss function when the input image is an unlabeled image, train the vascular lesion analysis model based on the reconstruction loss function, and return to the input unit; a second training unit is used to obtain a reconstruction loss function, a lesion type loss function, and a lesion severity loss function when the input image is a labeled image, determine whether the reconstruction loss function, the lesion type loss function, and the lesion severity loss function simultaneously satisfy a target condition, and if they do not simultaneously satisfy the target condition, then proceed to the input unit; if they simultaneously satisfy the target condition, then the trained vascular lesion analysis model is obtained; the target condition is that the reconstruction loss function is less than a preset reconstruction threshold, the lesion type loss function is less than a preset classification threshold, and the lesion severity loss function is less than a preset severity threshold.

[0153] Thus, the technical solution provided in this application can train the vascular lesion analysis model by using a large number of unlabeled images and a small number of labeled images as input images in a single-image alternating input manner. The trained vascular lesion analysis model can determine the lesion type and degree of vascular lesions in vascular images. This training method, which does not require a large number of labeled images, can reduce the manual annotation operation and reduce the workload of doctors. At the same time, the alternating training with unlabeled and labeled images, as well as the training method that simultaneously satisfies the target conditions based on the reconstruction loss function, lesion type loss function, and lesion degree loss function, also ensures the accuracy of vascular lesion determination.

[0154] Please see Figure 7 , Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 7 As shown, the electronic device 700 includes a processor 710, a memory 720, and a bus 730.

[0155] The memory 720 stores machine-readable instructions executable by the processor 710. When the electronic device 700 is running, the processor 710 communicates with the memory 720 via the bus 730. When the machine-readable instructions are executed by the processor 710, they can perform the operations described above. Figure 1 as well as Figure 2 The steps of the method for determining vascular lesions in the illustrated method embodiment can be found in the method embodiment for specific implementation methods, which will not be repeated here.

[0156] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can perform the above-described actions. Figure 1 as well as Figure 2 The steps of the method for determining vascular lesions in the illustrated method embodiment can be found in the method embodiment for specific implementation methods, which will not be repeated here.

[0157] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0158] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the shown or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0159] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0160] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0161] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0162] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The scope of protection of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method of determining a vascular lesion, characterized by, The determination method comprises: obtaining a blood vessel image; inputting the blood vessel image into a blood vessel lesion analysis model to output a lesion type result and a lesion degree result; wherein the blood vessel lesion analysis model is obtained through the following steps: (A) obtaining a plurality of unmarked images and a plurality of marked images, wherein the unmarked images are of a first preset threshold, the marked images are of a second preset threshold, the marked images are pre-labeled with lesion types and lesion degrees, and the first preset threshold is a preset multiple of the second preset threshold; (B) inputting the unmarked images and the marked images into the blood vessel lesion analysis model in a single-alternating manner; (C) if the input image is an unmarked image, a reconstruction loss function is obtained, the blood vessel lesion analysis model is trained according to the reconstruction loss function, and step (B) is performed again; (D) if the input image is a marked image, a reconstruction loss function, a lesion type loss function and a lesion degree loss function are obtained, it is determined whether the reconstruction loss function, the lesion type loss function and the lesion degree loss function satisfy a target condition at the same time, if not, step (B) is performed again, and if so, the trained blood vessel lesion analysis model is obtained; the target condition is that the reconstruction loss function is less than a preset reconstruction threshold, the lesion type loss function is less than a preset classification threshold, and the lesion degree loss function is less than a preset degree threshold; the blood vessel lesion analysis model comprises a feature extraction module and an image reconstruction module; the reconstruction loss function is obtained through the following steps: inputting the input image into the feature extraction module to extract features, obtaining a feature image corresponding to the feature extraction module; inputting the feature image into the image reconstruction module to reconstruct the image, obtaining a target reconstruction image output by the image reconstruction module; comparing the target reconstruction image with the input image to obtain a reconstruction loss function; the feature extraction module comprises a plurality of extraction network structures, each extraction network structure comprising N convolution layers and M pooling layers; the step of inputting the input image into the feature extraction module to extract features and obtaining a feature image corresponding to the feature extraction module comprises: inputting the input image into a first extraction network structure of the feature extraction module, taking the first extraction network structure as a current extraction network structure, extracting feature information of the input image through the N convolution layers of the current extraction network structure, and then condensing the extracted feature information through the M pooling layers of the current extraction network structure to obtain a feature image corresponding to the current extraction network structure; updating a next extraction network structure of the first extraction network structure to a current extraction network structure, inputting a feature image corresponding to the previous extraction network structure into the current extraction network structure as an input image, and continuing to obtain a feature image corresponding to the current extraction network structure until the last extraction network structure in the feature extraction module is reached, and obtaining a feature image corresponding to each extraction network structure in the feature extraction module; The blood vessel lesion analysis model further comprises a lesion analysis module; the lesion analysis module comprises a lesion classifier and a lesion degree regressor; if the input image is a labeled image, the lesion type loss function and the lesion degree loss function are obtained by the following steps: The feature images corresponding to each extraction network structure in the feature extraction module are input into the lesion analysis module for superposition and fusion to obtain a target feature image; The target feature image is input into the lesion classifier for classification to obtain a lesion type result, and the lesion type loss function is obtained according to the lesion type result and the pre-labeled lesion type in the labeled image; The target feature image is input into the lesion degree regressor for lesion degree prediction to obtain a lesion degree result, and the lesion degree loss function is obtained according to the lesion degree result and the pre-labeled lesion degree in the labeled image.

2. The determination method according to claim 1, characterized in that, The image reconstruction module comprises the same number of reconstruction network structures as the feature extraction module, and each reconstruction network structure comprises N convolutional layers and M deconvolutional layers; the step of inputting the feature images into the image reconstruction module for image reconstruction to obtain the target reconstruction image output by the image reconstruction module comprises: The feature images corresponding to each extraction network structure in the feature extraction module are input into the corresponding reconstruction network structure of the reconstruction network structure; The first reconstruction network structure of the image reconstruction module is taken as a current reconstruction network structure, and the feature image corresponding to the last extraction network structure of the feature extraction module is extracted by the N convolutional layers of the current reconstruction network structure and then is up-sampled by the M deconvolutional layers of the current reconstruction network structure to obtain a reconstruction image corresponding to the current reconstruction network structure; The next reconstruction network structure of the first reconstruction network structure is updated to be the current reconstruction network structure, and the reconstruction image corresponding to the previous reconstruction network structure and the feature image corresponding to the corresponding extraction network structure in the feature extraction module are input into the current reconstruction network structure to continue to obtain the reconstruction image corresponding to the current reconstruction network structure, until the last reconstruction network structure in the image reconstruction module is reached to obtain the target reconstruction image output by the last reconstruction network structure of the image reconstruction module.

3. The determination method according to claim 1, characterized in that, The number of the unlabeled images and the number of the labeled images are both pre-set thresholds; The data processing is performed by the following steps: For each labeled image and each unlabeled image, the labeled image and the unlabeled image are resampled according to a pre-set spatial resolution to obtain labeled images and unlabeled images with uniform spatial resolution; The labeled images and the unlabeled images with uniform spatial resolution are unified in image brightness according to a pre-set brightness range to obtain labeled images and unlabeled images with uniform image brightness; The image brightness-unified marked image and the unmarked image are subjected to standard score processing to obtain the unmarked image with a first preset threshold of data processing and the marked image with a second preset threshold of a pre-labeled lesion type and lesion degree.

4. The determination method according to claim 1, characterized in that, The step of inputting the blood vessel image into the blood vessel lesion analysis model and outputting the lesion type result and the lesion degree result comprises: The blood vessel image is input into the blood vessel lesion analysis model, and feature extraction is performed through a feature extraction module in the blood vessel lesion analysis model to obtain multiple feature images of the blood vessel image; The multiple feature images of the blood vessel image are subjected to lesion analysis through a lesion analysis module in the blood vessel lesion analysis model to obtain a lesion type result output by a lesion classifier of the lesion analysis module and a lesion degree result output by a lesion degree regressor of the lesion analysis module.

5. An apparatus for determining a vascular lesion, characterized by The determining device comprises: an acquisition module configured to acquire a blood vessel image; an output module configured to input the blood vessel image into a blood vessel lesion analysis model and output a lesion type result and a lesion degree result; a training module configured to train the blood vessel lesion analysis model; the training module comprises an acquisition unit, an input unit, a first training unit and a second training unit; the acquisition unit is configured to acquire a first preset threshold of unmarked images and a second preset threshold of marked images pre-labeled with a lesion type and a lesion degree; wherein the first preset threshold is a preset multiple of the second preset threshold; the input unit is configured to input the unmarked images and the marked images as input images into the blood vessel lesion analysis model in a single-alternating input manner; the first training unit is configured to obtain a reconstruction loss function when the input image is an unmarked image, train the blood vessel lesion analysis model according to the reconstruction loss function, and return to the input unit; the second training unit is configured to obtain a reconstruction loss function, a lesion type loss function and a lesion degree loss function when the input image is a marked image, determine whether the reconstruction loss function, the lesion type loss function and the lesion degree loss function simultaneously satisfy a target condition, if not, enter the input unit, and if so, obtain the trained blood vessel lesion analysis model; the target condition is that the reconstruction loss function is less than a preset reconstruction threshold, the lesion type loss function is less than a preset classification threshold, and the lesion degree loss function is less than a preset degree threshold; the blood vessel lesion analysis model comprises a feature extraction module and an image reconstruction module; when obtaining the reconstruction loss function, the first training unit and the second training unit are specifically configured to: input the input image into the feature extraction module for feature extraction to obtain a feature image corresponding to the feature extraction module; input the feature image into the image reconstruction module for image reconstruction to obtain a target reconstruction image output by the image reconstruction module; and the blood vessel lesion analysis model comprises a feature extraction module and an image reconstruction module; when obtaining the reconstruction loss function, the first training unit and the second training unit are specifically configured to: input the input image into the feature extraction module for feature extraction to obtain a feature image corresponding to the feature extraction module; input the feature image into the image reconstruction module for image reconstruction to obtain a target reconstruction image output by the image reconstruction module. The target reconstructed image and the input image are compared to obtain a reconstruction loss function; The feature extraction module includes a plurality of extraction network structures, each of which includes N convolutional layers and M pooling layers; the blood vessel lesion analysis model includes a feature extraction module and an image reconstruction module; when the first training unit and the second training unit are used to input the input image to the feature extraction module for feature extraction to obtain the feature image corresponding to the feature extraction module, the first training unit and the second training unit are specifically used to: input the input image into the first extraction network structure of the feature extraction module, and take the first extraction network structure as the current extraction network structure; after extracting the feature information of the input image through the N convolutional layers of the current extraction network structure, the extracted feature information is concentrated through the M pooling layers of the current extraction network structure to obtain the feature image corresponding to the current extraction network structure; the next extraction network structure of the first extraction network structure is updated to the current extraction network structure, and the feature image corresponding to the previous extraction network structure is input to the current extraction network structure as an input image to continue to obtain the feature image corresponding to the current extraction network structure, until the last extraction network structure in the feature extraction module is traversed to obtain the feature image corresponding to each extraction network structure in the feature extraction module; The blood vessel lesion analysis model further includes a lesion analysis module; the lesion analysis module includes a lesion classifier and a lesion degree regressor; when the second training unit is used to obtain a lesion type loss function and a lesion degree loss function, the second training unit is specifically used to: input the feature image corresponding to each extraction network structure in the feature extraction module into the lesion analysis module for superimposed fusion to obtain a target feature image; input the target feature image into the lesion classifier for classification to obtain a lesion type result, and obtain a lesion type loss function according to the lesion type result and the pre-labeled lesion type in the labeled image; input the target feature image into the lesion degree regressor for lesion degree prediction to obtain a lesion degree result, and obtain a lesion degree loss function according to the lesion degree result and the pre-labeled lesion degree in the labeled image.

6. An electronic device, comprising: comprise: a processor, a memory and a bus, the memory storing machine readable instructions executable by the processor, when the electronic device is running, the processor and the memory communicate through the bus, the machine readable instructions are executed by the processor to execute the steps of the blood vessel lesion determination method according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, the computer program is executed by the processor to execute the steps of the blood vessel lesion determination method according to any one of claims 1 to 4.

Images