Automatic classification system for predicting cardiovascular and cerebrovascular events

Abstract

The invention relates to the technical field of image recognition, in particular to an automatic classification system for predicting cardiovascular and cerebrovascular events, which comprises the following steps of: segmenting carotid plaques and tracking the carotid plaques frame by frame to obtain preliminary plaque segmentation results; fusing and predicting to obtain a fusion segmentation result; extracting quantitative features of the carotid plaque; and performing prediction according to the fusion segmentation result and the quantitative features, and outputting a classification probability result corresponding to the cardiovascular and cerebrovascular event. According to the method, the problem of poor single-modal recognition effect is effectively solved, the accuracy of cardiovascular and cerebrovascular event prediction is remarkably improved through technical means such as multi-modal data fusion and feature guide attention, and powerful support is provided for clinical diagnosis and risk assessment. A carotid artery plaque part in common ultrasonic data is tracked and enhanced frame by frame through a plaque tracking module in advance, so that a subsequent network can capture plaque related semantic features more easily. And image quantitative features are extracted and input into a classification module, so that more accurate classification prediction is realized.

Description

Technical Field

[0001] This invention relates to the field of multimodal data fusion image recognition technology, specifically to an automatic classification system for predicting cardiovascular and cerebrovascular events. Background Technology

[0002] Cardiovascular and cerebrovascular diseases are among the leading causes of death and disability worldwide. Rupture and detachment of atherosclerotic plaques in the carotid arteries are a major cause of coronary heart disease, stroke, and other cardiovascular and cerebrovascular diseases, seriously threatening human health. As a major artery connecting the heart and brain, the atherosclerotic plaques in the carotid artery can reflect the development trend of atherosclerosis in major arteries throughout the body. Therefore, early detection, diagnosis, and risk assessment of carotid atherosclerotic plaques are of significant practical and clinical value for developing personalized treatment plans and reducing the incidence and mortality of cardiovascular and cerebrovascular events.

[0003] Ultrasound imaging, with its advantages of being non-invasive, real-time, convenient, and low-cost, is the primary medical imaging method for monitoring and diagnosing carotid atherosclerosis. Because ultrasound videos contain richer information about carotid plaques and blood vessels, reflecting real-time changes in carotid plaques, ultrasound physicians typically analyze and diagnose based on ultrasound videos. In clinical applications, ultrasound physicians usually combine grayscale ultrasound examination with contrast-enhanced ultrasound to comprehensively assess plaque stability and more fully evaluate the risk of cardiovascular and cerebrovascular events. Driven by medical informatics and high-performance computing platforms, deep learning, with its powerful feature learning capabilities and classification performance, has been applied to the intelligent diagnosis of carotid plaques. Existing technologies mainly suffer from the following shortcomings: first, they can only process static images and cannot effectively utilize the spatiotemporal contextual information of ultrasound videos; second, most algorithms rely only on single-modal information, lacking comprehensive analysis of multimodal plaque features, resulting in limited prediction accuracy. However, current research methods have some problems. First, current research mainly focuses on static images, ignoring the spatiotemporal contextual information provided by ultrasound. Secondly, most algorithms only utilize information from a single modality and lack multimodal plaque features, making it difficult to conduct a comprehensive analysis of risks and affecting the accuracy of cardiovascular and cerebrovascular event prediction. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, an automatic classification system for predicting cardiovascular and cerebrovascular events is provided.

[0005] The specific technical solution is as follows: An automated classification system for predicting cardiovascular and cerebrovascular events, comprising: The acquisition module collects normal ultrasound data, enhanced ultrasound data, and network feature maps from the patient for subsequent fusion analysis. A patch tracking module, which is connected to the acquisition module; The plaque tracking module segments the carotid artery plaques in the normal ultrasound data and tracks them frame by frame to obtain preliminary plaque segmentation results. Its main function is to provide preliminary plaque segmentation results and enhance the contextual information of subsequent modules. A multimodal fusion prediction module, which is connected to both the acquisition module and the patch tracking module; The multimodal fusion prediction module fuses the preliminary patch segmentation result, the enhanced ultrasound data, and the network feature map, and predicts the fused segmentation result. A feature-guided attention module, which is connected to the acquisition module; The feature-guided attention module extracts quantitative features of carotid plaques from the normal ultrasound data and the enhanced ultrasound data; A multi-layer perception prediction module, which is connected to the multi-modal fusion prediction module and the feature-guided attention module respectively; The multilayer perception prediction module makes predictions based on the quantized features on the basis of the fusion segmentation results, so as to output classification probability results corresponding to different cardiovascular and cerebrovascular events.

[0006] On the other hand, the acquisition module includes: An image acquisition module is connected to an external ultrasound acquisition device and acquires the normal ultrasound data and the enhanced ultrasound data, respectively. A network extraction module is connected to the image acquisition module; The network extraction module extracts the network feature map from the typical ultrasound data based on a pre-trained extraction model.

[0007] On the other hand, a patch edge attention module has been added to the extraction model; The plaque edge attention module automatically tracks and segments carotid artery plaques to enhance edge information.

[0008] On the other hand, the patch tracking module includes: A dual attention detector, wherein the dual attention detector uses two attention channels to segment the carotid plaque region and the vascular region respectively for the input enhanced ultrasound data; Tracker, the tracker being connected to the dual attention detector; The tracker tracks the carotid plaque region and the vascular region detected in each of the multiple historical frames of the enhanced ultrasound data and marks them to obtain the preliminary segmentation result of the plaque.

[0009] On the other hand, the dual attention detector is implemented based on the Attention U-Net network; The tracker is implemented based on MSTUnet.

[0010] On the other hand, the multimodal fusion prediction module includes: A multimodal fusion module fuses the preliminary patch segmentation results, the enhanced ultrasound data, and the network feature map to obtain fused features; A block embedding module, which is connected to the multimodal fusion module; The block embedding module embeds the fusion features to obtain an embedding map; The Trasnformer network is connected to the block embedding module; The Trasnformer network predicts the fused segmentation result based on the attention mechanism of the embedded graph.

[0011] On the other hand, the Trasnformer network includes: A time-dimensional prediction unit, which captures global context information of the embedded graph in chronological order to obtain a first prediction feature; A spatial dimension prediction unit, which is connected to the time dimension prediction unit; The spatial dimension prediction unit extracts semantic features from the first prediction feature to obtain the second prediction feature; The first normalization module is connected to the spatial dimension prediction unit; The first normalization module performs normalization processing based on the second predicted feature to obtain the first normalized feature; MLP sensor, the MLP sensor being connected to the second normalization module; The MLP perceptron generates the fused segmentation result based on the second normalized feature.

[0012] On the other hand, the time dimension prediction unit and the spatial dimension prediction unit use the same prediction unit and adjust the attention mechanism to achieve this; The prediction unit includes: The second normalization module performs normalization processing on the input image to obtain the second normalized features; A multi-head convolutional attention module, wherein the multi-head convolutional attention module is connected to the second normalization module; The multi-head convolutional attention module extracts the target feature from the second normalized feature based on an attention mechanism; The addition module is connected to the multi-head convolutional attention module; The addition module adds the target features to the original input of the prediction unit to obtain the output result.

[0013] On the other hand, the feature-guided attention module includes: A mask adding module, wherein the mask adding module adds masks to the normal ultrasound data and the enhanced ultrasound data respectively to form mask images; Feature calculation module, which is connected to the mask addition module; The feature calculation module calculates image features based on the mask image; The image features include intensity features, texture features, wavelet features, patch features, contrast features, clinical features, and classification indicators; The U-test module is connected to the feature calculation module; The U-test module calculates the importance of the image features and outputs a predetermined number of image features with high significance as the quantified features.

[0014] On the other hand, the multilayer sensing prediction module includes: The MLP output head predicts the classification result corresponding to the cardiovascular and cerebrovascular event based on the quantized features and the fusion segmentation result. The probability calculation module is connected to the MLP output head; The probability calculation module generates the classification probability result based on the classification result.

[0015] The above technical solution has the following advantages or beneficial effects: To address the issue of poor performance in carotid plaque recognition using a single modality in existing technologies, this solution introduces multimodal ultrasound video data for fusion recognition. Furthermore, to improve recognition accuracy, a plaque tracking module pre-tracks and enhances the carotid plaque portion of the typical ultrasound data frame by frame, enabling the subsequent network to more easily capture semantic features associated with the plaque. Additionally, by extracting quantified features from the image and inputting them into the classification module, the classification module can achieve more accurate prediction results during the classification process. Attached Figure Description

[0016] Embodiments of the invention will be described more fully with reference to the accompanying drawings. However, the drawings are for illustration and explanation only and do not constitute a limitation on the scope of the invention.

[0017] Figure 1 This is an overall schematic diagram of an embodiment of the present invention; Figure 2 This is a schematic diagram of the acquisition module in an embodiment of the present invention; Figure 3 This is a schematic diagram of the patch tracking module in an embodiment of the present invention; Figure 4 This is a schematic diagram of the multimodal fusion prediction module in an embodiment of the present invention; Figure 5 This is a schematic diagram of the feature-guided attention module in an embodiment of the present invention; Figure 6 This is a schematic diagram of the multilayer sensing prediction module in an embodiment of the present invention; Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0020] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0021] This invention includes: An automated classification system for predicting cardiovascular and cerebrovascular events, such as Figure 1 As shown, it includes: Acquisition module 1 acquires conventional ultrasound data, enhanced ultrasound data, and network feature maps from the patient. Patch tracking module 2, which is connected to acquisition module 1; The plaque tracking module 2 segments carotid artery plaques in typical ultrasound data and tracks them frame by frame to obtain preliminary plaque segmentation results; Multimodal fusion prediction module 3 is connected to acquisition module 1 and patch tracking module 2 respectively; The multimodal fusion prediction module 3 fuses the preliminary patch segmentation results, enhanced ultrasound data, and network feature maps, and predicts the fused segmentation results. Feature-guided attention module 4 is connected to acquisition module 1; Feature-guided attention module 4 extracts quantitative features of carotid plaques from normal ultrasound data and enhanced ultrasound data; Multi-layer perception prediction module 5, which is connected to multi-modal fusion prediction module 3 and feature-guided attention module 4 respectively; The multilayer perception prediction module 5 makes predictions based on quantitative features on the basis of the fused segmentation results, so as to output the classification probability results corresponding to different cardiovascular and cerebrovascular events.

[0022] Specifically, to address the problem that the single-modal approach in existing technologies is not effective in identifying carotid plaques, this solution introduces multimodal ultrasound video data for fusion recognition.

[0023] The multimodal ultrasound video data includes standard ultrasound data, contrast-enhanced ultrasound data, and network feature maps. Standard ultrasound data (Ultrasound Data) consists of grayscale ultrasound images acquired through routine scanning procedures. Contrast-enhanced ultrasound data (CEUS) is obtained by intravenously injecting microbubble contrast agent after a routine ultrasound examination, providing a better representation of carotid artery perfusion characteristics. Network feature maps are generated by pre-processing a network from the standard ultrasound data to extract target boundaries and perfusion features. This effectively identifies irregular and ulcerated fibrous caps in the carotid artery, providing a basis for subsequent network extraction.

[0024] To achieve better extraction of plaques from typical ultrasound data, a plaque tracking module 2 is used to segment carotid plaques in the typical ultrasound data. The segmented target results are tracked in multiple video frames of the ultrasound data, thereby providing global contextual features for subsequent models and improving the continuity of segmentation.

[0025] Subsequently, the multimodal fusion prediction module 3 fuses the preliminary plaque segmentation results, enhanced ultrasound data, and network feature maps, and predicts the fused segmentation results to achieve effective extraction of the fibrous cap portion of carotid plaques, perfusion features, calcified areas, and ulceration areas. Furthermore, considering the diagnostic characteristics of different ultrasound doctors during the scanning process, the features of image parameters are quantified for the acquired images themselves, and highly significant quantified features are extracted to provide the multilayer perception prediction module 5 as a basis for prediction.

[0026] Finally, based on the fused segmentation results, predictions are made using quantified features to output classification probability results corresponding to different cardiovascular and cerebrovascular events. In one embodiment, such as Figure 2 As shown, the acquisition module 1 includes: Image acquisition module 11 is connected to an external ultrasound acquisition device and acquires normal ultrasound data and enhanced ultrasound data respectively. Network extraction module 12, which is connected to image acquisition module 11; The network extraction module 12 extracts network feature maps from ordinary ultrasound data based on a pre-trained extraction model.

[0027] Specifically, in order to achieve better multimodal data extraction results, in this embodiment, the image acquisition module 11 is first connected to an external ultrasound acquisition device to acquire normal ultrasound data and enhanced ultrasound data respectively.

[0028] Standard ultrasound data (SUS) consists of grayscale ultrasound images acquired through routine scanning procedures. Contrast-enhanced ultrasound data (CEUS) is obtained by intravenously injecting microbubble contrast agent after a standard ultrasound examination, providing a better representation of carotid artery perfusion characteristics. The network feature map is a feature map obtained by pre-processing an extraction network from standard ultrasound data to extract target boundaries and perfusion features. It can effectively identify irregular and ulcerated fibrous caps in the carotid artery, providing a basis for subsequent network extraction.

[0029] The network feature map is a feature map obtained by pre-extracting target boundaries and perfusion features based on conventional ultrasound data using an extraction model. It can effectively identify irregular and ulcerated fibrous caps in the carotid artery and provide a basis for subsequent network extraction.

[0030] The extraction model is a multimodal ultrasound video analysis network BP-Net based on target boundaries and perfusion features.

[0031] The model includes a patch edge attention module. The plaque edge attention module automatically tracks and segments carotid artery plaques to enhance edge information.

[0032] The plaque edge attention module automatically extracts the carotid artery plaque region and fibrous cap, mimicking the steps taken by a doctor, thus overcoming the limitations of traditional attention mechanisms. Furthermore, this module improves computational efficiency and processing speed.

[0033] The above structure enables better pre-extraction of patchy parts in video frames, thereby clearly marking target boundaries and infusion features in the network feature map.

[0034] In one embodiment, such as Figure 3 As shown, the patch tracking module 2 includes: Dual attention detector 21 uses two attention channels to segment the carotid plaque region and the vascular region respectively for the input enhanced ultrasound data; Tracker 22, tracker 22 is connected to dual attention detector 21; Tracker 22 tracks the carotid plaque region and vascular region detected in each of the multiple historical frames of enhanced ultrasound data and marks them to obtain preliminary plaque segmentation results.

[0035] Specifically, in order to facilitate the subsequent network to better capture contextual features during the extraction of multimodal features, in this embodiment, an AMPTS framework is constructed in the patch tracking module 2 to automatically and accurately track and segment multiple patches in the ultrasound video to enhance the features of this part.

[0036] Among them, the dual attention detector 21 uses two attention channels to segment the carotid plaque region and the vascular region respectively for the input enhanced ultrasound data. Each attention channel changes the segmentation object by adjusting the attention mechanism during training.

[0037] Subsequently, in order to effectively track segmented objects on multiple frames, tracker 22 is also sent to track the carotid plaque region and vascular region detected in each frame of multiple historical frames of enhanced ultrasound data and label them to obtain preliminary plaque segmentation results, thereby capturing short-range and long-range contextual features and temporal information to improve the accuracy of tracking and segmentation.

[0038] Specifically, the dual attention detector is implemented based on the Attention U-Net network; the tracker is implemented based on MSTUnet.

[0039] The dual attention detector is a multi-task deep learning method based on Attention U-Net, which can simultaneously segment carotid plaques and blood vessels. The tracker, MSTUnet, combines a multi-stream similarity learning network with a convolutional neural network for medical image segmentation. It can utilize features from historical frames to better track the target in the current frame and is robust to significant changes in appearance.

[0040] In one embodiment, such as Figure 4 As shown, the multimodal fusion prediction module 3 includes: The multimodal fusion module 31 fuses the preliminary patch segmentation results, enhanced ultrasound data, and network feature maps to obtain fused features; Block embedding module 32, which is connected to multimodal fusion module 31; Block embedding module 32 embeds the fused features to obtain an embedding map; Trasnformer network 33, Trasnformer network 33 connection block embedding module 32; The Trasnformer network 33 predicts the fused segmentation result of the embedded graph based on the attention mechanism.

[0041] Specifically, to achieve better recognition results, in this embodiment, a multimodal fusion module 31 is first used to fuse the preliminary patch segmentation results, enhanced ultrasound data, and network feature maps to obtain fused features. Specifically, multimodal fusion can integrate features of the preliminary patch segmentation results generated based on ordinary ultrasound data, including patch size, shape, texture, and patch edges, as well as perfusion and ulceration information of enhanced ultrasound data and further learn higher-level features based on the BP-Net network feature maps. Integrating multimodal information effectively enhances useful information while eliminating redundant information.

[0042] Subsequently, the fused features are embedded by the block embedding module 32 to obtain an embedding map. Then, the fused segmentation result is obtained by predicting the embedding map based on the attention mechanism using the Trasnformer network 33. This captures complex semantic features and global contextual information, further improving the accuracy and robustness of the network. In one embodiment, such as Figure 4 As shown, the Trasnformer network 33 includes: The time dimension prediction unit 331 captures global context information of the embedded graph in chronological order to obtain the first prediction feature; Spatial dimension prediction unit 332, which is connected to time dimension prediction unit 331; The spatial dimension prediction unit 332 extracts semantic features from the first prediction feature to obtain the second prediction feature; The first normalization module 333 is connected to the spatial dimension prediction unit 332; The first normalization module 333 performs normalization processing based on the second predicted feature to obtain the first normalized feature; MLP perceptron 334, MLP perceptron 334 is connected to the second normalization module 333; The MLP perceptron 334 generates a fused segmentation result based on the second normalized feature.

[0043] Specifically, in order to better capture the contextual information in continuous image frames and obtain information on carotid plaques, perfusion features, calcification areas, and ulcers in continuous ultrasound videos, this embodiment adds a temporal dimension prediction unit 331 to the Trasnformer network 33 to capture global contextual information, and a spatial dimension prediction unit 332 to capture complex semantic features. This enables effective tracking of target regions that appear continuously in the temporal dimension in ultrasound data, as well as accurate segmentation of target regions on a single frame image. Finally, the corresponding classification results are output through the MLP perceptron 334.

[0044] In one embodiment, the time dimension prediction unit 331 and the spatial dimension prediction unit 332 use the same prediction unit and adjust the attention mechanism to achieve this. The prediction unit includes: The second normalization module 3301 performs normalization processing on the input image to obtain the second normalized feature; Multi-head convolutional attention module 3302, which is connected to the second normalization module 3301; The multi-head convolutional attention module 3302 extracts the target features from the second normalized features based on the attention mechanism; The addition module 3303 is connected to the multi-head convolutional attention module 3302; The addition module 3303 adds the target features to the original input of the prediction unit to obtain the output result.

[0045] Specifically, to achieve better feature recognition results, this embodiment introduces a multi-head convolutional attention module 3302 into the prediction unit, which is embedded in the Transformer network to extract target features based on the attention mechanism. This combines the advantages of convolution and Transformer, capturing complex semantic features and global contextual information, further improving the network's accuracy and robustness.

[0046] In one embodiment, such as Figure 5 As shown, the feature-guided attention module 4 includes: Masking module 41 adds masks to both normal ultrasound data and enhanced ultrasound data to form mask images; Feature calculation module 42, which is connected to mask addition module 41; Feature calculation module 42 calculates image features based on the mask image; Image features include intensity features, texture features, wavelet features, patch features, contrast features, clinical features, and classification indicators; U-test module 43, U-test module 43 is connected to feature calculation module 42; The U-test module 43 calculates the importance of image features and outputs a predetermined number of image features with high significance as quantification features.

[0047] Specifically, to achieve better feature selection, in this embodiment, the feature calculation module 42 first calculates image features based on the mask image, extracting radiomics features, plaque grayscale features, plaque angiography features, clinical features, and classification results of carotid plaque calcification from both regular and enhanced ultrasound data. Then, the U-test module 43 performs a Mann-Whitney U-test on the image features to select valuable features and integrates them into the deep learning model, further improving prediction accuracy.

[0048] In one embodiment, such as Figure 6 As shown, the multilayer sensing prediction module 5 includes: MLP output head 51 predicts the classification results corresponding to cardiovascular and cerebrovascular events based on quantized features and fusion segmentation results; Probability calculation module 52, which is connected to MLP output head 51; The probability calculation module generates classification probability results based on the classification results.

[0049] To comprehensively evaluate and analyze the effectiveness of the DK-Net network, three parts of experiments were conducted. First, the DK-Net network was compared with eight state-of-the-art classification methods, including three Transformer-based methods, two convolutional neural network-based methods, and three traditional machine learning-based methods. DK-Net achieved 86.94% ACC, 0.9446 AUC, 98.44% Sens, and 74.36% Spec. Second, a series of ablation experiments were conducted to analyze the necessity of each modality in DK-Net, including the US modality, CEUS modality, network feature map modality, and IF features. Finally, a series of ablation experiments were conducted to analyze the performance of each module of DK-Net. The clinical feature-guided attention module, multimodal fusion module, and multi-head convolutional attention module all improved the model performance to varying degrees, thus validating the effectiveness of the proposed modules. The results show that DK-Net can achieve accurate prediction of cardiovascular and cerebrovascular events, assist clinicians in diagnosis, and help with risk stratification and personalized treatment strategies.

[0050] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made based on the description and illustrations of the present invention should be included within the protection scope of the present invention.

Claims

1. An automatic classification system for predicting cardiovascular and cerebrovascular events, characterized in that, include: The acquisition module acquires conventional ultrasound data, enhanced ultrasound data, and network feature maps from the patient. A patch tracking module, which is connected to the acquisition module; The plaque tracking module segments the carotid plaques in the normal ultrasound data and tracks them frame by frame to obtain preliminary plaque segmentation results. A multimodal fusion prediction module, which is connected to both the acquisition module and the patch tracking module; The multimodal fusion prediction module fuses the preliminary patch segmentation result, the enhanced ultrasound data, and the network feature map, and predicts the fused segmentation result. A feature-guided attention module, which is connected to the acquisition module; The feature-guided attention module extracts quantitative features of carotid plaques from the normal ultrasound data and the enhanced ultrasound data; A multi-layer perception prediction module, which is connected to the multi-modal fusion prediction module and the feature-guided attention module respectively; The multilayer perception prediction module makes predictions based on the quantized features on the basis of the fusion segmentation results, so as to output classification probability results corresponding to different cardiovascular and cerebrovascular events.

2. The automatic classification system according to claim 1, characterized in that, The acquisition module includes: An image acquisition module is connected to an external ultrasound acquisition device and acquires the normal ultrasound data and the enhanced ultrasound data, respectively. A network extraction module is connected to the image acquisition module; The network extraction module uses a pre-trained extraction model to extract the network feature map from the typical ultrasound data.

3. The automatic classification system according to claim 2, characterized in that, The extraction model includes a patch edge attention module. The plaque edge attention module automatically tracks and segments carotid artery plaques, thereby enhancing edge information.

4. The automatic classification system according to claim 1, characterized in that, The patch tracking module includes: A dual attention detector, wherein the dual attention detector uses two attention channels to segment the carotid plaque region and the vascular region respectively for the input enhanced ultrasound data; Tracker, the tracker being connected to the dual attention detector; The tracker tracks and labels the carotid plaque region and the vascular region detected in multiple historical frames of the enhanced ultrasound data, respectively, to obtain preliminary plaque segmentation results.

5. The automatic classification system according to claim 4, characterized in that, The dual attention detector is implemented using the Attention U-Net network; the tracker is implemented using MSTUnet.

6. The automatic classification system according to claim 1, characterized in that, The multimodal fusion prediction module includes: A multimodal fusion module fuses the preliminary patch segmentation results, the enhanced ultrasound data, and the network feature map to obtain fused features; A block embedding module, which is connected to the multimodal fusion module; The block embedding module embeds the fusion features to obtain an embedding map; The Trasnformer network is connected to the block embedding module; The Trasnformer network predicts the fused segmentation result based on the attention mechanism of the embedded graph.

7. The automatic classification system according to claim 6, characterized in that, The Trasnformer network includes: A time-dimensional prediction unit, which captures global context information of the embedded graph in chronological order to obtain a first prediction feature; A spatial dimension prediction unit, which is connected to the time dimension prediction unit; The spatial dimension prediction unit extracts semantic features from the first prediction feature to obtain the second prediction feature; The first normalization module is connected to the spatial dimension prediction unit; The first normalization module performs normalization processing based on the second predicted feature to obtain the first normalized feature; MLP sensor, the MLP sensor being connected to the second normalization module; The MLP perceptron generates the fused segmentation result based on the second normalized feature.

8. The automatic classification system according to claim 7, characterized in that, The time dimension prediction unit and the spatial dimension prediction unit use the same prediction unit and adjust the attention mechanism to achieve this. The prediction unit includes: The second normalization module performs normalization processing on the input image to obtain the second normalized features; A multi-head convolutional attention module, wherein the multi-head convolutional attention module is connected to the second normalization module; The multi-head convolutional attention module extracts the target feature from the second normalized feature based on an attention mechanism; The addition module is connected to the multi-head convolutional attention module; The addition module adds the target features to the original input of the prediction unit to obtain the output result.

9. The automatic classification system according to claim 1, characterized in that, The feature-guided attention module includes: A mask adding module, wherein the mask adding module adds masks to the normal ultrasound data and the enhanced ultrasound data respectively to form mask images; Feature calculation module, which is connected to the mask addition module; The feature calculation module calculates image features based on the mask image; The image features include intensity features, texture features, wavelet features, patch features, contrast features, clinical features, and classification indicators; The U-test module is connected to the feature calculation module; The U-test module calculates the importance of the image features and outputs a predetermined number of image features with high significance as the quantified features.

10. The automatic classification system according to claim 1, characterized in that, The multilayer sensing prediction module includes an MLP output head and a probability calculation module; The MLP output head predicts the classification results corresponding to cardiovascular and cerebrovascular events based on quantization features and fusion segmentation results; The probability calculation module is connected to the MLP output head, and the probability calculation module generates the classification probability result based on the classification result.

Images