CN118071688A - Real-time cerebral angiography quality assessment method - Google Patents

Abstract

The present invention relates to the field of medical technology, and discloses a method for real-time angiography quality assessment during angiography, which can assist doctors in angiography quality control. The real-time cerebral angiography quality assessment method eliminates the influence of background on quality assessment. At the same time, in view of the segmentation difficulties of DSA angiography, the network introduces a window-based local-global self-attention mechanism with reference to the global feature extraction advantage of Transformer. While ensuring linear complexity, it pays more attention to the correlation between the various parts of the entire vascular structure, effectively improving the segmentation accuracy. In addition, a feature aggregation module is designed to filter encoder features based on the attention method, making feature aggregation more efficient; at the same time, the method combines angiography clinical quality evaluation indicators, determines the quality classification indicators based on the angiography image quality classification data set, uses an automated method to locate image quality control points, and uses random forests for quality assessment.

Description

A real-time cerebral angiography quality assessment method

Technical Field

The invention relates to the field of medical technology, and in particular to a real-time cerebral angiography quality assessment method.

Background technique

Cerebrovascular disease (CVD) is a disease that causes brain tissue damage due to intracranial blood circulation disorders. It has become one of the major diseases that endanger human health and life.

At present, many studies have introduced deep learning image processing methods for automatic quality assessment of medical images. Quality assessment is usually performed on the basis of image processing, such as image quality assessment based on object detection or segmentation of the main body. Due to the complexity of contrast images, a two-stage quality assessment method of image processing plus quality classification is adopted.

For the processing of vascular images, segmentation networks such as U-Net are usually used to segment blood vessels and background, which has good segmentation performance. The imaging time of DSA angiography is usually only about 12 seconds. Quality assessment during the process requires real-time image analysis, and insufficient computing resources need to be considered when deployed to medical equipment. Therefore, as an important basis for quality assessment, the real-time and lightweight nature of the segmentation algorithm must be considered. However, there are some special segmentation difficulties in cerebral angiography images, such as the simultaneous presence of large-diameter arteries and small-diameter capillaries, uneven distribution of contrast agents, uneven X-ray exposure, etc., and the usual vascular segmentation algorithms have problems such as large number of parameters, high amount of calculation, and slow reasoning speed. On the one hand, the existing cerebral angiography segmentation methods designed for the characteristics of DSA have good segmentation accuracy, but there are relatively few studies on real-time lightweight segmentation of cerebral angiography, and the balance between accuracy and real-time nature of cerebral vascular DSA segmentation has not been well solved. On the other hand, the current lightweight methods of medical image segmentation, such as lightweight feature extraction networks, attention mechanisms, and adding various feature fusion modules to replace network depth, have improved the reasoning speed to a certain extent. However, due to insufficient feature extraction and other reasons, it is difficult to ensure the quality of segmentation for difficult segmentation problems such as cerebrovascular segmentation. Based on the above problems, a segmentation network was designed for the characteristics of cerebrovascular DSA, and the real-time and accuracy of the segmentation network were ensured through specially designed feature processing modules and lightweight methods.

At present, the image quality assessment method uses manual feature extraction combined with machine learning algorithms for classification. Manually extracted features are more intuitive. In recent years, the classification method based on deep learning has emerged. It performs quality classification by autonomously learning sample features and has good generalization performance.

On the one hand, the sample size of the established angiography image dataset is small and is not suitable for deep learning classification methods that require a large amount of training data. On the other hand, there are relatively clear quality assessment indicators for cerebral angiography quality assessment, and the manual feature extraction method can simulate the doctor's evaluation process and accurately perform quality assessment. Therefore, the method of extracting features and performing machine learning classification is more suitable for angiography image quality assessment, and an automated feature extraction method is designed to achieve automated quality assessment.

Summary of the invention

1. Technical issues to be solved

In view of the shortcomings of the prior art, the present invention provides a real-time cerebral angiography quality assessment method, which has the advantages of small amount of calculation and accurate extraction, and solves the problem of large amount of calculation in image evaluation and the need to improve extraction efficiency and accuracy.

(II) Technical solution

In order to achieve the above-mentioned purpose of small amount of calculation and accurate extraction, the present invention provides the following technical solution: a real-time cerebral angiography quality assessment method, comprising the following steps:

Step S1, the collected internal carotid artery anteroposterior images are respectively subjected to vessel segmentation and quality classification annotation, angiography image segmentation and classification data sets are established, and training sets and test sets are divided. The specific steps are as follows;

Step S1.1, collecting 110 internal carotid artery anteroposterior images and establishing angiography image dataset;

Step S1.2, the segmentation and classification data sets share the above-mentioned angiography image data, and the segmentation annotation is manually annotated, and the annotated areas are the vascular trunks and branches;

Step S1.3, the quality classification categories include qualified and unqualified. Unqualified quality mainly includes the cases of too high or too low contrast concentration, abnormal vascular structure, foreign body artifacts and motion artifacts;

Step S2: Input the image into the segmentation model for training, using the full-image training method. The segmentation model is improved based on U-Net, including a lightweight feature extraction backbone, a local-global self-attention mechanism, and a feature aggregation module;

Step S2.1, lightweight feature extraction backbone: First, for angiography quality control evaluation, only the main trunk and main branches of the blood vessels need to be segmented, and full-image training can help the network learn the vascular structure; second, the full-image training method does not require pre- and post-processing. When the amount of calculation increases significantly after high-resolution image input, deep separable convolution is used to replace traditional convolution to reduce the amount of calculation and achieve a lightweight backbone;

Step S2.2, local-global self-attention mechanism: The encoder and decoder introduce a self-attention mechanism, which can fully extract the vascular structure features based on the full image input;

Step S2.3, feature aggregation module: weight the encoder features through spatial attention to retain valid information;

Step S3, inputting the classification data set and the corresponding index calculation results into the quality classification model for training, obtaining the final classification model, designing the quality assessment method according to the clinical quality control index, clinically judging the quality of angiography mainly through the grayscale size of vascular development, the uniformity of contrast agent, the integrity of vascular structure, and the abnormal shape of blood vessels, designing appropriate quality control indicators for these dimensions, selecting the quality control area in the main vascular trunk, and then determining the specific quality control points for calculating the quality control index, including the following steps;

Step S3.1, the quality control area includes the C2-C3 segment and the C6-C7 segment of the internal carotid artery trunk, which can reflect the quality problems caused by abnormal contrast agent concentration and uniformity; the YOLOv7 lightweight target detection model is used to automatically locate the quality control area, and a detection data set is established for training;

Step S3.2, the quality control point is the maximum inscribed circle of the blood vessel in the quality control area, and the positioning method is: extract the blood vessel contour, select the midpoint of the height of the quality control area as the y-axis coordinate value of the center of the quality control point, and use the maximum inscribed circle radius method to find the quality control point in the quality control area;

Step S3.3, select the overall area and grayscale mean of the blood vessel, the grayscale mean and variance of the quality control points, and the number of quality control areas detected as quality control indicators. Among them, the overall area and grayscale mean of the blood vessel are the number of pixel points and grayscale mean of the segmented blood vessel area, the grayscale mean and variance of the quality control point are the grayscale mean and variance of the quality control point-inscribed circle-inner pixel, and the number of quality control areas detected is the number of quality control areas detected by the target detection model;

Step S3.4, the quality classification model adopts the random forest algorithm. The random forest is composed of multiple decision trees. The final prediction result is generated by voting the results of multiple decision trees to avoid randomness and make the classification result more stable. The random forest training and test data sets use the quality classification data sets and the corresponding indicator data. The indicator data comes from the calculation results of the image indicators of the classification data set.

Preferably, in step S1, the training set and the test set are divided in a ratio of 7:3, the training set is used for model training, and the test set is used to evaluate model performance. The segmentation evaluation indicators include accuracy, precision, sensitivity, and parameter quantity Params, and computing quantity FLOPs. The classification evaluation indicators include accuracy, precision, sensitivity, and the average inference time is used to evaluate the real-time performance of the overall system.

Preferably, in step S2, the segmentation model is based on a U-Net encoder-decoder architecture, and the encoder feature extraction uses the MBConv deep separable convolution block in EfficientNet to reduce the number of model parameters;

Among them, a local-global self-attention module L-GBlock is introduced after each MBConv. Through the multi-scale window sizes of different feature layers, local to global feature information modeling is realized. When the feature map size is reduced, the image receptive field is expanded from local to global, which can fully extract the local coherence information of blood vessels and the global structural information.

The local-global self-attention module is symmetrically distributed on the encoder and decoder ends. A feature aggregation module FAM is introduced at the connection between the encoder and decoder features. The encoder features are summed with the upsampled features and the spatial attention matrix is calculated. Finally, the original encoder features are weighted for attention and concatenated with the upsampled features.

Preferably, in step S2.1, the first 7 layers of EfficientNet are used as the feature extraction backbone to reduce network parameters and computational complexity.

Preferably, in step S2.2, Transformer first uses a self-attention mechanism to model global feature interactions, and the self-attention mechanism for two-dimensional images will increase quadratically with the image resolution. In order to achieve global information interaction and keep the computational complexity linear, a window-based self-attention mechanism is used;

First, the features are divided into P×P windows of the same size, and each window is subjected to FFN self-attention calculation to obtain the attention map. The original features are weighted to obtain the global attention features, and the window size hyperparameter is selected;

The number of windows is set to 16, 8, 4, 2, and 1 respectively, with the feature resolution decreasing as the network depth increases. In the shallow network, local self-attention is performed, while in the deep network, self-attention is performed on the entire feature.

A dynamic local-to-global self-attention mechanism is implemented. A symmetric self-attention module is introduced on the decoder side, which can ensure the integrity of the vascular structure and local coherence while reconstructing the resolution. The number of windows on the decoder side is set to 2, 4, 8, and 16 respectively.

Preferably, in step S2.3, the dimensions of the two channels are first reduced and summed, and then adaptive convolution is performed and normalized using the Sigmoid function to obtain a spatial probability map; finally, the features of the attention mechanism are combined with the upsampling features, and subsequent self-attention calculations are performed; the feature aggregation modules all use 1×1 convolution to transform the channel dimension, which is computationally small and lightweight.

(III) Beneficial effects

Compared with the prior art, the present invention provides a real-time cerebral angiography quality assessment method, which has the following beneficial effects:

This real-time cerebral angiography quality assessment method uses an automated method to locate image quality control points and is suitable for deep learning classification methods that require a large amount of training data. By introducing a lightweight real-time segmentation network based on U-Net, the collected samples use separable convolutions with smaller computational complexity. At the same time, in view of the segmentation difficulties of DSA angiography, a window-based local-global self-attention mechanism is introduced. While ensuring linear complexity, it pays more attention to the correlation between the various parts of the entire vascular structure, effectively improving the segmentation accuracy. In addition, the feature aggregation module filters the encoder features based on the attention method, making feature aggregation more efficient, improving the accuracy of feature extraction for image acquisition, and optimizing and improving the evaluation effect of cerebral angiography quality.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of a cerebral angiography quality assessment method proposed by the present invention;

FIG. 2 is a schematic diagram of the structure of a lightweight blood vessel segmentation model according to the present invention.

Detailed ways

The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

Please refer to Figures 1-2: A real-time cerebral angiography quality assessment method, comprising the following steps:

Step S1, collect the anteroposterior images of the internal carotid artery, perform vascular segmentation and quality classification annotation, establish angiography image segmentation and classification data set, and divide the training set and test set. In step S1, the training set and test set are divided according to the ratio of 7:3. The training set is used for model training, and the test set is used to evaluate model performance. The segmentation evaluation indicators include accuracy, precision, sensitivity, as well as the number of parameters and the amount of FLOPs. The classification evaluation indicators include accuracy, precision, sensitivity, and the average inference time is used to evaluate the real-time performance of the overall system. The specific steps are as follows;

Step S1.1, collecting 110 internal carotid artery anteroposterior images and establishing angiography image dataset;

Step S1.2, the segmentation and classification data sets share the above-mentioned angiography image data, and the segmentation annotation is manually annotated, and the annotated areas are the vascular trunks and branches;

Step S1.3, the quality classification categories include qualified and unqualified. Unqualified quality mainly includes the cases of too high or too low contrast concentration, abnormal vascular structure, foreign body artifacts and motion artifacts;

Step S2, please refer to Figure 2, input the image into the segmentation model for training, and adopt the whole-image training method, including lightweight feature extraction backbone, local-global self-attention mechanism and feature aggregation module; the segmentation model is designed based on the U-Net encoder-decoder architecture, and the encoder feature extraction adopts the MBConv deep separable convolution block in EfficientNet, which significantly reduces the number of model parameters. In order to make up for the insufficient feature extraction of the lightweight convolution block, a local-global self-attention module (L-GBlock) is introduced after each MBConv, and the local to global feature information modeling is realized through the multi-scale window size of different feature layers. As the size of the feature map decreases, the receptive field expands from local to global, and the local coherence information of the blood vessels and the global structural information can be fully extracted. The local-global self-attention module is symmetrically distributed on the encoder and decoder ends. A feature aggregation module (FAM) is introduced at the connection between the encoder and decoder features, and the encoder features are summed with the upsampled features and the spatial attention matrix is calculated. Finally, the original encoder features are weighted for attention and spliced with the upsampled features. This module can extract information that is beneficial to segmentation from the original image features, avoiding the information redundancy and feature differences of direct short connections.

Step S2.1, lightweight feature extraction backbone: First, for angiography quality control evaluation, it is only necessary to segment the main trunk and main branches of the blood vessels, and full-image training can help the network learn the vascular structure; second, the full-image training method does not require pre- and post-processing. When the amount of calculation increases significantly after high-resolution image input, deep separable convolution is used to replace traditional convolution to reduce the amount of calculation and achieve a lightweight backbone; the first 7 layers of EfficientNet are used as the feature extraction backbone to reduce network parameters and the amount of calculation.

Step S2.2, local-global self-attention mechanism: The encoder and decoder introduce a self-attention mechanism, which can fully extract the vascular structure features based on the full image input; Transformer first uses the self-attention mechanism to model global feature interactions, and the self-attention mechanism for two-dimensional images will increase quadratically with the image resolution; in order to achieve global information interaction and keep the computational complexity linear, a window-based self-attention mechanism is used;

First, the features are divided into P×P windows of the same size, and each window is subjected to FFN self-attention calculation to obtain the attention map. The original features are weighted to obtain the global attention features, and the window size hyperparameter is selected;

The number of windows is set to 16, 8, 4, 2, and 1 respectively, with the feature resolution decreasing as the network depth increases. In the shallow network, local self-attention is performed, while in the deep network, self-attention is performed on the entire feature.

A dynamic local-to-global self-attention mechanism is implemented. A symmetric self-attention module is introduced on the decoder side, which can ensure the integrity of the vascular structure and local coherence while reconstructing the resolution. The number of windows on the decoder side is set to 2, 4, 8, and 16 respectively.

Step S2.3, feature aggregation module: weight the encoder features through spatial attention to retain valid information; first reduce the dimension of the two channels and sum them, then use adaptive convolution and sigmoid function to normalize and obtain a spatial probability map; finally, combine the features of the attention mechanism with the upsampled features, and perform subsequent self-attention calculations; the feature aggregation modules all use 1×1 convolution for channel dimension transformation, which is computationally small and lightweight.

Step S3, input the classification data set and the corresponding index calculation results into the quality classification model for training, and obtain the final classification model. The quality assessment method is designed according to the clinical quality control index. The clinical judgment of the angiography quality is mainly based on the grayscale size of vascular development, the uniformity of the contrast agent, the integrity of the vascular structure, and the abnormal shape of the vascular. Appropriate quality control indicators are designed for these dimensions, and the quality control area is selected in the main trunk of the blood vessel, and then the specific quality control points are determined to calculate the quality control indicators, including the following steps;

Step S3.1, the quality control area includes the C2-C3 segment and the C6-C7 segment of the internal carotid artery trunk, which can reflect the quality problems caused by abnormal contrast agent concentration and uniformity; the YOLOv7 lightweight target detection model is used to automatically locate the quality control area, and a detection data set is established for training;

Step S3.2, the quality control point is the maximum inscribed circle of the blood vessel in the quality control area, and the positioning method is: extract the blood vessel contour, select the midpoint of the height of the quality control area as the y-axis coordinate value of the center of the quality control point, and use the maximum inscribed circle radius method to find the quality control point in the quality control area;

Step S3.3, select the overall area and grayscale mean of the blood vessel, the grayscale mean and variance of the quality control points, and the number of quality control areas detected as quality control indicators. Among them, the overall area and grayscale mean of the blood vessel are the number of pixel points and grayscale mean of the segmented blood vessel area, the grayscale mean and variance of the quality control point are the grayscale mean and variance of the quality control point-inscribed circle-inner pixel, and the number of quality control areas detected is the number of quality control areas detected by the target detection model;

Step S3.4, the quality classification model adopts the random forest algorithm. The random forest is composed of multiple decision trees. The final prediction result is generated by voting the results of multiple decision trees to avoid randomness and make the classification result more stable. The random forest training and test data sets use the quality classification data sets and the corresponding indicator data. The indicator data comes from the calculation results of the image indicators of the classification data set.

Experimental example: The experiment was conducted on the test set of the self-built cerebral angiography quality classification dataset. The classification accuracy of the proposed cerebral angiography quality control method based on vascular segmentation reached 84.6%, and the average execution time was 0.87s. A comparative experiment was conducted on the test set of the self-built cerebral angiography segmentation dataset. The designed lightweight segmentation network segmentation accuracy reached 98.0%, which is almost equivalent to the performance of the advanced segmentation network. The segmentation network parameter volume is only 2.89MB, which is much smaller than the current advanced segmentation model. The inference FPS is 27, achieving real-time segmentation performance. The ablation experiment proves that the proposed local-global self-attention mechanism and feature fusion module have a significant effect on improving network performance.

The beneficial effects of the present invention are:

This real-time cerebral angiography quality assessment method uses an automated method to locate image quality control points and is suitable for deep learning classification methods that require a large amount of training data. By introducing a lightweight real-time segmentation network based on U-Net, the collected samples use separable convolutions with smaller computational complexity. At the same time, in view of the segmentation difficulties of DSA angiography, a window-based local-global self-attention mechanism is introduced. While ensuring linear complexity, it pays more attention to the correlation between the various parts of the entire vascular structure, effectively improving the segmentation accuracy. In addition, the feature aggregation module filters the encoder features based on the attention method, making feature aggregation more efficient, improving the accuracy of feature extraction for image acquisition, and optimizing and improving the evaluation effect of cerebral angiography quality.

Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims (6)

1. A real-time cerebral angiography quality assessment method, characterized in that it comprises the following steps: Step S1, the collected internal carotid artery anteroposterior images are respectively subjected to vessel segmentation and quality classification annotation, angiography image segmentation and classification data sets are established, and training sets and test sets are divided. The specific steps are as follows; Step S1.1, collecting 110 internal carotid artery anteroposterior images and establishing angiography image dataset; Step S1.2, the segmentation and classification data sets share the above-mentioned angiography image data, and the segmentation annotation is manually annotated, and the annotated areas are the vascular trunks and branches; Step S1.3, the quality classification categories include qualified and unqualified. Unqualified quality mainly includes the cases of too high or too low contrast concentration, abnormal vascular structure, foreign body artifacts and motion artifacts; Step S2: Input the image into the segmentation model for training, using the full-image training method. The segmentation model is improved based on U-Net, including a lightweight feature extraction backbone, a local-global self-attention mechanism, and a feature aggregation module; Step S2.1, lightweight feature extraction backbone: First, for angiography quality control evaluation, only the main trunk and main branches of the blood vessels need to be segmented, and full-image training can help the network learn the vascular structure; second, the full-image training method does not require pre- and post-processing. When the amount of calculation increases significantly after high-resolution image input, deep separable convolution is used to replace traditional convolution to reduce the amount of calculation and achieve a lightweight backbone; Step S2.2, local-global self-attention mechanism: The encoder and decoder introduce a self-attention mechanism, which can fully extract the vascular structure features based on the full image input; Step S2.3, feature aggregation module: weight the encoder features through spatial attention to retain valid information; Step S3, inputting the classification data set and the corresponding index calculation results into the quality classification model for training, obtaining the final classification model, designing the quality assessment method according to the clinical quality control index, clinically judging the quality of angiography mainly through the grayscale size of vascular development, the uniformity of contrast agent, the integrity of vascular structure, and the abnormal shape of blood vessels, designing appropriate quality control indicators for these dimensions, selecting the quality control area in the main vascular trunk, and then determining the specific quality control points for calculating the quality control index, including the following steps; Step S3.1, the quality control area includes the C2-C3 segment and the C6-C7 segment of the internal carotid artery trunk, which can reflect the quality problems caused by abnormal contrast agent concentration and uniformity; the YOLOv7 lightweight target detection model is used to automatically locate the quality control area, and a detection data set is established for training; Step S3.2, the quality control point is the maximum inscribed circle of the blood vessel in the quality control area, and the positioning method is: extract the blood vessel contour, select the midpoint of the height of the quality control area as the y-axis coordinate value of the center of the quality control point, and use the maximum inscribed circle radius method to find the quality control point in the quality control area; Step S3.3, select the overall area and grayscale mean of the blood vessel, the grayscale mean and variance of the quality control points, and the number of quality control areas detected as quality control indicators. Among them, the overall area and grayscale mean of the blood vessel are the number of pixel points and grayscale mean of the segmented blood vessel area, the grayscale mean and variance of the quality control point are the grayscale mean and variance of the quality control point-inscribed circle-inner pixel, and the number of quality control areas detected is the number of quality control areas detected by the target detection model; Step S3.4, the quality classification model adopts the random forest algorithm. The random forest is composed of multiple decision trees. The final prediction result is generated by voting the results of multiple decision trees to avoid randomness and make the classification result more stable. The random forest training and test data sets use the quality classification data sets and the corresponding indicator data. The indicator data comes from the calculation results of the image indicators of the classification data set. 2. A real-time cerebral angiography quality assessment method according to claim 1, characterized in that in step S1, the training set and the test set are divided according to a ratio of 7:3, the training set is used for model training, and the test set is used to evaluate model performance, the segmentation evaluation indicators include accuracy, precision, sensitivity, as well as parameters Params, and calculation amount FLOPs, the classification evaluation indicators include accuracy, precision, sensitivity, and the average inference time is used to evaluate the real-time performance of the overall system. 3. A real-time cerebral angiography quality assessment method according to claim 1, characterized in that in step S2, the segmentation model is based on a U-Net encoder-decoder architecture, and the encoder feature extraction adopts the MBConv deep separable convolution block in EfficientNet to reduce the number of model parameters; Among them, a local-global self-attention module L-GBlock is introduced after each MBConv. Through the multi-scale window sizes of different feature layers, local to global feature information modeling is realized. When the size of the feature map is reduced, the receptive field is expanded from local to global, which can fully extract the local coherence information of the blood vessels and the global structural information; The local-global self-attention module is symmetrically distributed on the encoder and decoder ends. A feature aggregation module FAM is introduced at the connection between the encoder and decoder features. The encoder features are summed with the upsampled features and the spatial attention matrix is calculated. Finally, the original encoder features are weighted for attention and concatenated with the upsampled features. 4. A real-time cerebral angiography quality assessment method according to claim 1, characterized in that, in step S2.1, the first 7 layers of EfficientNet are used as the feature extraction backbone to reduce network parameters and calculation amount. 5. A real-time cerebral angiography quality assessment method according to claim 1, characterized in that, in step S2.2, Transformer first uses a self-attention mechanism to model global feature interaction, and the self-attention mechanism for two-dimensional images will increase quadratically with the image resolution; in order to achieve global information interaction and keep the computational complexity linear, a window-based self-attention mechanism is used; First, the features are divided into P×P windows of the same size, and each window is subjected to FFN self-attention calculation to obtain the attention map. The original features are weighted to obtain the global attention features, and the window size hyperparameter is selected; The number of windows is set to 16, 8, 4, 2, and 1 respectively, with the feature resolution decreasing as the network depth increases. In the shallow network, local self-attention is performed, while in the deep network, self-attention is performed on the entire feature. A dynamic local-to-global self-attention mechanism is implemented. A symmetric self-attention module is introduced on the decoder side, which can ensure the integrity of the vascular structure and local coherence while reconstructing the resolution. The number of windows on the decoder side is set to 2, 4, 8, and 16 respectively. 6. A real-time cerebral angiography quality assessment method according to claim 1, characterized in that, in step S2.3, firstly, the dimensions of the two channels are reduced and summed, and then the spatial probability map is obtained by adaptive convolution and normalization using the Sigmoid function; finally, the features of the attention mechanism are combined with the up-sampled features, and subsequent self-attention calculations are performed; the feature aggregation modules all use 1×1 convolution to transform the channel dimension, which has a small amount of calculation and is lightweight.