Aortic dissection prediction method and system based on U-Net double attention mechanism
The aortic dissection prediction method using the U-Net dual attention mechanism solves the problems of segmentation accuracy and efficiency, achieving high-precision aortic and intimal flap segmentation and DeBakey classification, meeting clinical needs and improving the efficiency and accuracy of diagnosis and treatment.
Patent Information
- Application Number
- CN202512044254.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-31
- Publication Date
- 2026-04-10
AI Technical Summary
Existing segmentation models cannot meet clinical needs in terms of segmentation accuracy and efficiency when dealing with aortic dissections with complex anatomical structures.
A method for predicting aortic dissection based on the U-Net dual attention mechanism is adopted. By introducing the channel attention module SEBlock and the spatial attention module CBAM, and combining multi-level deep supervision branches, fine segmentation of the aorta and intimal flap is performed, and knowledge distillation and pruning strategies are used to optimize the network structure.
It significantly improves the segmentation accuracy of the aorta and intimal flap, shortens the diagnosis time, meets the rapid diagnosis needs in emergency scenarios, and achieves accurate DeBakey classification, thereby improving treatment outcomes and patient recovery rates.
Smart Images
Figure CN121837233A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of medical image processing, in particular to an aortic dissection prediction method and system based on a U-Net double attention mechanism and a storage medium. BACKGROUND
[0002] As a cardiovascular emergency with rapid onset and high mortality, timely and accurate diagnosis is crucial for saving the lives of patients. Currently, in the field of aortic dissection diagnosis, traditional diagnosis methods mainly rely on manual processing of medical images, especially CTA (CT angiography) images. This approach has many drawbacks. In terms of image processing, traditional methods require a lot of time, and doctors need to manually identify key structures such as the aorta and intimal flap. In complex medical images, this process is not only tedious but also highly susceptible to subjective factors. Differences in judgment standards and experience among different doctors can lead to inconsistencies in diagnosis results. In Medical Imaging (Editor: Jin Zhengyu, People's Medical Publishing House, published in 2018), it is mentioned that it can take several minutes or even longer to manually process a CTA image, which undoubtedly delays the best treatment opportunity for patients in the emergency scene. In Cardiovascular Disease Imaging Diagnosis (Editor: Guo Qiyong, People's Medical Publishing House, published in 2017), it is pointed out that traditional segmentation models have low segmentation accuracy when dealing with complex anatomical structures of aortic dissection, which cannot meet clinical needs. In addition, traditional methods also face difficulties in distinguishing between true and false lumens of aortic dissection. In some patient cases, the boundary between true and false lumens is not obvious, and traditional segmentation methods rely on accurate judgment of true and false lumens, which makes the segmentation process complex and the results unreliable in such cases. When processing high-resolution CTA images, traditional techniques are difficult to balance speed and accuracy, and cannot meet the real-time requirements of clinical practice. SUMMARY
[0003] The technical problem to be solved by the present application is:
[0004] The existing segmentation model cannot meet the clinical needs in terms of segmentation accuracy and efficiency when dealing with complex anatomical structures of aortic dissection.
[0005] The technical solution adopted by the present application to solve the above technical problems is:
[0006] The present application provides an aortic dissection prediction method based on a U-Net double attention mechanism, comprising the following steps:
[0007] Step S100, slice preprocessing of the CTA data to be detected;
[0008] Step S200, input the preprocessed slices into the first segmentation network for aortic segmentation to obtain aortic region slice images and aortic categories;
[0009] Step S300: Input the aortic region slice image into the second segmentation network based on U-Net for intimal flap segmentation. Its encoder includes multiple feature extraction layers with progressive downsampling. Each feature extraction layer introduces a channel attention module SEBlock. The SEBlock is used to obtain the global response information of each channel and learn the weight of each channel by performing global pooling on the current feature map, so as to adaptively recalibrate the features of different channels.
[0010] Its decoder includes multiple progressively upsampled feature recovery layers and fuses features from the corresponding layers of the encoder to obtain feature maps. The decoder introduces a spatial attention module CBAM, which generates a spatial attention heatmap by performing pooling and convolution operations on the feature map in the spatial dimension, enabling the network to focus on the subtle differences at the edge of the endometrial flap during the decoding stage.
[0011] The decoder has deep supervision branches at different levels, which connect multiple intermediate feature layers of different resolutions and are used to apply auxiliary supervision constraints to the intermediate feature outputs.
[0012] Step S400: Calculate the positive percentage of aortic dissection for each aortic category, and perform DeBakey classification of aortic dissection based on the extent of dissection involvement.
[0013] Furthermore, step S200 also includes: sorting the CTA slices according to the order of their corresponding slice layers in the CTA image to be detected along the stacking direction to form a slice sequence; using a sliding window to filter the junctions of different segments of the aorta, dividing the slice sequence into slice sub-sequences of different aortic categories, classifying each slice sub-sequence into the corresponding segment according to the number and shape of the aortic regions detected in the slice, and removing slice images in each slice sub-sequence that do not belong to the aortic category.
[0014] Furthermore, the first segmentation network is based on the U-Net network, and residual blocks are constructed at each block to prevent gradient vanishing and expand the receptive field.
[0015] Furthermore, the deep supervision branch connects at least three intermediate feature layers with different resolutions. Each deep supervision branch includes a convolutional mapping layer and an upsampling layer, so that the output of the intermediate feature layer is mapped to a spatial scale consistent with the final segmentation result.
[0016] Furthermore, in the training phase of the second segmentation network in step 300, a knowledge distillation method is adopted. A dual-domain feature distillation framework is constructed to address the characteristics of significant differences in fine-grained features and blurred boundaries of the endometrial flap structure in CTA images. This framework includes two distillation paths: semantic domain distillation and structural domain distillation. This allows the student network to simultaneously learn the high-level semantic representation and low-level structural details of the teacher model during the training process.
[0017] Furthermore, step S300 also includes: for the channel features output by each convolutional layer in the second segmentation network decoder and encoder, firstly, the normalized weight response value during the training phase is calculated as an importance index of the channel contribution; then, the convolutional kernels corresponding to the convolutional channels with importance scores lower than a preset threshold are pruned and permanently removed from the network structure so that the channel is no longer calculated during the inference phase.
[0018] Furthermore, the aortic category mentioned in step 400 includes three anatomical segments: the ascending aorta, the aortic arch, and the descending aorta. Based on the distribution of the intimal flap and false lumen in each segment, the segment involved in the dissection is determined, providing a basis for the DeBakey classification.
[0019] Furthermore, the method also includes: performing three-dimensional vascular reconstruction based on the aortic segmentation map to generate a three-dimensional model of the aorta, and further performing three-dimensional vascular reconstruction again based on the intimal flap segmentation results to further improve the three-dimensional structural model of aortic dissection.
[0020] This invention provides an aortic dissection prediction system based on the U-Net dual attention mechanism. The system has a program module corresponding to the steps of any of the above-described technical solutions, and executes the steps in the above-described aortic dissection prediction method based on the U-Net dual attention mechanism when running.
[0021] The present invention provides a computer-readable storage medium storing a computer program configured to, when invoked by a processor, implement the steps of the aortic dissection prediction method based on the U-Net dual attention mechanism described in any of the above technical solutions.
[0022] Compared with the prior art, the beneficial effects of the present invention are:
[0023] The second segmentation network of this invention uses a channel attention module to filter features during the encoding stage and a spatial attention module to locate features during the decoding stage. It also combines multi-level deep supervision branches to constrain intermediate features, forming a feature learning mechanism that coordinates channel, spatial, and hierarchical dimensions. The segmentation model based on the dual attention fusion mechanism can significantly improve the segmentation accuracy of key anatomical structures such as the aorta and intimal flaps in complex medical imaging backgrounds.
[0024] (1) Improved accuracy: Through multi-attention mechanism and deep supervision training, the segmentation accuracy of the aorta and intimal flap is significantly improved when processing medical images, with a sensitivity of 0.923 and F1 = 0.920. Compared with traditional methods, it can more accurately define the aortic region and intimal flap, providing a reliable basis for subsequent classification and treatment planning.
[0025] (2) Improved efficiency: By combining pruning and distillation strategies, inference efficiency is improved while maintaining accuracy. It only takes 8 minutes from image upload to PDF archiving. This greatly shortens the diagnosis time, meets the need for rapid diagnosis in emergency scenarios, and saves patients valuable treatment time. Compared with traditional manual processing methods, the processing efficiency is improved by several times.
[0026] (3) Precise Classification: The dissection score is calculated, and the key anatomical coordinates of the DeBakey classification are automatically marked to achieve precise classification. This assists doctors in developing personalized treatment plans that better suit the individual patient's condition, thereby improving treatment effectiveness and the patient's chances of recovery. In clinical trials, the classification accuracy of this invention reached 88.1%, which is significantly higher than the classification accuracy of traditional methods.
[0027] (4) Multidimensional analysis support: Three-dimensional aortic vascular reconstruction fully restores the three-dimensional anatomical structure of aortic dissection from multiple dimensions, stages and modalities, avoiding the limitations of a single perspective, providing doctors with more comprehensive and intuitive information, which helps doctors to understand the condition more deeply and make more accurate diagnoses and decisions. Attached Figure Description
[0028] Figure 1 This is a schematic diagram of the aortic dissection prediction system based on the U-Net dual attention mechanism in an embodiment of the present invention;
[0029] Figure 2 The flowchart below illustrates the aortic dissection prediction method based on the U-Net dual attention mechanism in this invention. Figure 1 ;
[0030] Figure 3 The flowchart below illustrates the aortic dissection prediction method based on the U-Net dual attention mechanism in this invention. Figure 2 . Detailed Implementation
[0031] To enable those skilled in the art to better understand the present invention, exemplary embodiments or examples of the present invention will be described below in conjunction with the accompanying drawings. Obviously, the described embodiments or examples are merely some, not all, of the embodiments or examples of the present invention. All other embodiments or examples obtained by those skilled in the art based on the embodiments or examples of the present invention without inventive effort should fall within the scope of protection of the present invention.
[0032] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0033] Specific Implementation Plan 1: Combining Figures 1-3 As shown, this invention provides a method for predicting aortic dissection based on the U-Net dual attention mechanism, comprising the following steps:
[0034] Step S100: Perform slice preprocessing on the CTA data to be detected;
[0035] Step S200: Input the preprocessed slices into the first segmentation network to segment the aorta, and obtain the aortic region slice map and aortic category;
[0036] Step S300: Input the aortic region slice image into the second segmentation network based on U-Net for intimal flap segmentation. Its encoder includes multiple feature extraction layers with progressive downsampling. Each feature extraction layer introduces a channel attention module SEBlock. The SEBlock is used to obtain the global response information of each channel and learn the weight of each channel by performing global pooling on the current feature map, so as to adaptively recalibrate the features of different channels.
[0037] Its decoder includes multiple progressively upsampled feature recovery layers and fuses features from the corresponding layers of the encoder to obtain feature maps. The decoder introduces a spatial attention module CBAM, which generates a spatial attention heatmap by performing pooling and convolution operations on the feature map in the spatial dimension, enabling the network to focus on the subtle differences at the edge of the endometrial flap during the decoding stage.
[0038] The decoder has deep supervision branches at different levels, which connect multiple intermediate feature layers of different resolutions and are used to apply auxiliary supervision constraints to the intermediate feature outputs.
[0039] Step S400: Calculate the positive percentage of aortic dissection for each aortic category, and perform DeBakey classification of aortic dissection based on the extent of dissection involvement.
[0040] The second segmentation network disclosed in this implementation scheme adopts an improved encoder-decoder structure based on U-Net. It consists of an encoder, a decoder, skip connections, and multi-level supervisory branches, used for fine segmentation of the aorta and intimal flap region. While maintaining a symmetrical hierarchical structure, the encoder and decoder introduce a dual attention mechanism to enhance the network's ability to represent key anatomical features.
[0041] 1. Encoder Structure
[0042] The encoder comprises multiple progressively downsampled feature extraction layers. Each layer consists of convolution operations, normalization operations, and nonlinear activation functions, used to progressively extract multi-scale representations from low-level texture features to high-level semantic features. In each feature extraction layer of the encoder, a channel attention module (SEBlock) is introduced and positioned after the convolution operation and before the downsampling operation.
[0043] The SEBlock module performs global pooling on the current feature map to obtain global response information for each channel, and uses fully connected mapping to learn the weight coefficients of each channel, thereby adaptively recalibrating the features of different channels to enhance the contrast between the vessel wall and surrounding tissue. In this way, the encoder can strengthen the effective feature channels related to the vessel wall before feature compression, while suppressing redundant channels corresponding to surrounding tissue and background noise, improving the discriminative power of subsequent feature representations.
[0044] 2. Decoder Structure
[0045] The decoder is configured to correspond to the encoder layers, including multiple progressively upsampled feature recovery layers to gradually restore high-level semantic information to the original spatial resolution. During each decoding stage, the decoder fuses feature information from the corresponding encoder layer using skip connections to simultaneously preserve both spatial details and semantic information.
[0046] A spatial attention module (CBAM) is introduced into the decoder and placed on the feature map after skip connections are fused. The CBAM module generates a spatial attention heatmap by performing pooling and convolution operations on the feature map in the spatial dimension, assigning different weights to different spatial locations. This allows the network to focus on fine structural regions such as the edge of the endometrial flap and the tear during the decoding stage. This structural design avoids introducing spatial constraints prematurely before the features are fully fused, which helps improve the localization accuracy of small-scale anatomical structures.
[0047] 3. In-depth supervision
[0048] Deep supervision branches are further set at multiple different levels of the decoder. These branches are connected to multiple intermediate feature layers of different resolutions in the decoder, and are used to apply auxiliary supervision constraints to the intermediate feature outputs. Each deep supervision branch includes a convolutional mapping layer and an upsampling operation, enabling the intermediate layer outputs to be mapped to a spatial scale consistent with the final segmentation result.
[0049] By introducing a multi-level deep supervision structure into the decoder, the network can not only rely on the error signal of the final output layer during backpropagation, but also obtain effective gradient feedback from the intermediate layers. This enhances the model's ability to learn the detailed structure of the aortic boundary and intimal flap, and reduces the risk of information decay of fine-grained features in deep networks during training.
[0050] like Figures 1-3 As shown, this invention adopts a progressive framework of "coarse segmentation - fine segmentation":
[0051] The first stage (CNN1) involves inputting preprocessed 256×256 CTA slices and using a U-Net network for aortic segmentation. The network extracts and analyzes features from the input CTA image, outputs the aortic contour, and performs 3D vessel reconstruction based on this contour, thus initially constructing a 3D model of the aorta and providing a foundation for subsequent precise analysis.
[0052] The second stage (CNN2): The aortic region output from CNN1 is adjusted to 256×256 pixels and then input into the CNN2 network. In this stage, a dual attention mechanism (CBAM + SEBlock) is introduced. SEBlock learns channel weights through global pooling in the channel dimension, enhancing the contrast between the vessel wall and surrounding tissues; CBAM generates an attention heatmap through spatial convolution in the spatial dimension, accurately focusing on subtle differences at the edge of the intimal flap, achieving precise segmentation of the intimal flap. Then, three-dimensional vessel reconstruction is performed again to further improve the three-dimensional structural model of aortic dissection.
[0053] Specific Implementation Scheme Two: Step S200 further includes: sorting the CTA slices according to the order of their corresponding slice layers along the stacking direction in the CTA image to be detected, forming a slice sequence; to reduce errors caused by inaccurate segmentation (e.g., incorrectly detecting two aortic regions in a segment where only the descending aorta exists), a sliding window of size 5 slices is used to filter the boundaries between different aortic segments, dividing the slice sequence into slice sub-sequences of different aortic categories (ascending aorta segment, aortic arch segment, descending aorta segment). Based on the number and shape of the aortic regions detected within the slice, each slice sub-sequence is classified into the corresponding segment, and slice images that do not belong to the aortic category in each slice sub-sequence are removed. For each CTA volume, the dissection frequency of the ascending aorta and descending aorta is calculated separately, thereby providing data support for subsequent diagnosis. The rest of this implementation scheme is the same as Specific Implementation Scheme One.
[0054] Specific Implementation Scheme 3: The first segmentation network is based on the U-Net network, constructing residual blocks at each block to prevent gradient vanishing and expand the receptive field. This allows for better extraction of aortic location information and prevents other more circular structures from being identified as false positives. This implementation scheme is otherwise identical to Specific Implementation Scheme 2.
[0055] Specific Implementation Scheme 4: The deep supervision branch connects at least three intermediate feature layers with different resolutions. Each deep supervision branch includes a convolutional mapping layer and an upsampling layer, so that the output of the intermediate feature layers is mapped to a spatial scale consistent with the final segmentation result. This implementation scheme is otherwise the same as Specific Implementation Scheme 3.
[0056] Specific Implementation Scheme Five: In the training phase of the second segmentation network in step 300, a knowledge distillation method is adopted. A dual-domain feature distillation framework is constructed to address the significant differences in fine-grained features and the blurred boundaries of the endometrial flap structure in CTA images. This framework includes two distillation paths: semantic domain distillation and structural domain distillation. This allows the student network to simultaneously learn the high-level semantic representation and low-level structural details of the teacher model during training. All other aspects of this implementation scheme are the same as in Specific Implementation Scheme Four.
[0057] Specific Implementation Scheme Six: Step S300 further includes: For the channel features output by each convolutional layer in the second segmentation network decoder and encoder, firstly, calculating their normalized weight response values during the training phase as an importance indicator of channel contribution; then, pruning the convolutional kernels corresponding to convolutional channels with importance scores below a preset threshold, permanently removing them from the network structure so that these channels are no longer calculated during the inference phase. Removing redundant convolutional kernels through channel pruning reduces model size and improves inference efficiency. The rest of this implementation scheme is the same as Specific Implementation Scheme Five.
[0058] Specific Implementation Scheme Seven: Step 300 includes adjusting the pixel values of the aortic region slice image to a preset range. All other aspects of this implementation scheme are the same as Specific Implementation Scheme Six.
[0059] Specific Implementation Scheme Eight: The aortic categories mentioned in step 400 include three anatomical segments: the ascending aorta, the aortic arch, and the descending aorta. Based on the distribution of the intimal flap and false lumen in each segment, the affected segments of the dissection are determined, providing a basis for the DeBakey classification. All other aspects of this implementation scheme are the same as in Specific Implementation Scheme Seven.
[0060] Specific Implementation Scheme Nine: The method further includes: performing three-dimensional vascular reconstruction based on the aortic segmentation map to generate a three-dimensional model of the aorta; further performing three-dimensional vascular reconstruction again based on the intimal flap segmentation results to further improve the three-dimensional structural model of the aortic dissection. This supports interactive rotation and cross-sectional observation, intuitively presenting vascular morphology (such as stenosis, dilation, tortuosity, or occlusion), and assisting in the localization of anatomical abnormalities. It provides doctors with more intuitive and comprehensive anatomical structural information. Other aspects of this implementation scheme are the same as in Specific Implementation Scheme Eight.
[0061] This invention combines functional assessment with dissection scoring and the DeBakey classification standard to quantitatively evaluate the functional status of vascular bundles (such as the degree of blood flow obstruction and branch vessel occlusion), providing quantitative basis for clinical decision-making. It realizes a closed-loop process from image segmentation to three-dimensional visualization and then to quantitative assessment, significantly improving the efficiency and accuracy of aortic dissection diagnosis.
[0062] The aortic dissection prediction method (algorithm) based on the U-Net dual attention mechanism proposed in this invention is the underlying technical core of this invention, and various products can be derived based on the algorithm.
[0063] Based on the method proposed in this invention, an aortic dissection prediction system based on the U-Net dual attention mechanism is developed using a programming language. This system has program modules corresponding to the steps of the above technical solution, and executes the steps in the above aortic dissection prediction method based on the U-Net dual attention mechanism when running.
[0064] The developed system (software) computer program is stored on a computer-readable storage medium. This computer program is configured to implement the steps of the aortic dissection prediction method based on the U-Net dual attention mechanism when invoked by a processor. In other words, the invention is materialized on a carrier, becoming a computer program product.
[0065] Various implementations of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, application-specific integrated circuits (ASICs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include: implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.
[0066] The computational programs (also referred to as programs, software, software applications, or code) of this invention include machine instructions of a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or apparatus (e.g., disk, optical disk, memory, programmable logic device PLD) for providing machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.
[0067] The beneficial effects of the present invention will be described below with reference to specific embodiments.
[0068] Example 1
[0069] In this embodiment, CTA aortic dissection case images from Tongji Hospital in Wuhan and the Second Affiliated Hospital of Harbin Medical University were selected as the training and testing datasets. The aortic region and intimal flap region, annotated using the 3D Slicer software, were used as the ground truth labels. The images were divided into training and testing sets in an 8:2 ratio, and data augmentation techniques such as random blurring, contrast enhancement, and random noise were employed to increase sample diversity.
[0070] The model structure employs the dual-attention mechanism U-Net of this invention. The encoder includes a channel attention module (SEBlock) to weight features along the channel dimension, enhancing vessel wall features and suppressing background interference. The decoder includes a spatial attention module (CBAM) to generate attention heatmaps in the spatial dimension, highlighting the intimal flap edge and tear area. Furthermore, deep supervision branches are set at at least three different layers of the decoder to provide auxiliary supervision of the intermediate feature layer outputs, enhancing the network's ability to learn the aortic boundary and intimal flap details. During training, the AdamW optimizer is used with an initial learning rate of 0.001, and the model is trained end-to-end using a combination of Dice loss and binary cross-entropy loss.
[0071] After training, the model was applied to an independent test set for performance validation. Test results show that in pixel-level segmentation of the aorta and intimal flap, the proposed dual-attention mechanism and deep-supervised training method significantly improve model performance. With an IoU threshold of 0.6, the model's sensitivity reaches 0.923 and its F1 score reaches 0.920, both significantly outperforming traditional U-Net and segmentation models without attention mechanisms. This demonstrates that the multi-attention mechanism employed in this embodiment effectively enhances the model's ability to focus on key regions, while the deep-supervised strategy further strengthens the intermediate layers' ability to capture boundary information, enabling the model to maintain stable and accurate segmentation performance even in challenging scenarios such as complex vascular backgrounds, overlapping true and false lumens, and blurred intimal flap boundaries.
[0072] Example 3
[0073] In this embodiment, aortic dissection CTA images from the Second Affiliated Hospital of Harbin Medical University and Tongji Hospital of Huazhong University of Science and Technology were selected as validation data. All cases were independently interpreted by cardiovascular imaging experts and agreed upon as reference standards. First, the multi-attention segmentation network of this invention was used to obtain high-precision segmentation results of the aortic lumen, false lumen, true lumen, and intimal flap. Based on this, the inventors constructed an automated dissection scoring model, which formed a quantitative scoring system for the severity of dissection by measuring key indicators such as the false lumen extent, the ratio of true to false lumen area, and the location of intimal flap tear in the ascending aorta, aortic arch, and descending aorta regions.
[0074] Subsequently, key anatomical coordinates such as the origin of the ascending aorta, the opening of the brachiocephalic artery, the opening of the left subclavian artery, and the level of the diaphragm are located. The system automatically identifies DeBakey types I, II, and III aortic dissections by determining the distribution of the tear and the extent of the dissection in different segments of the aorta. The entire classification process requires no manual intervention, and the automatic classification results are compared with those obtained manually by experts. Experimental results show that the automated classification accuracy of this invention reaches 88.1%, significantly higher than traditional classification methods. The classification system of this invention maintains high stability even under challenging conditions such as complex false lumen ranges, unclear boundaries, and variations in the anatomical structure of the aorta arch.
[0075] As demonstrated in this embodiment, the automatic dissection scoring and DeBakey classification method proposed in this invention can not only perform rapid and accurate structured analysis of aortic dissection, but also provide more individualized and precise treatment strategies for clinical practice, improving treatment success rates and patient recovery chances, thus possessing significant clinical application value.
[0076] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.
Claims
1. A method for predicting aortic dissection based on U-Net dual attention mechanism, characterized in that, Includes the following steps: Step S100: Perform slice preprocessing on the CTA data to be detected; Step S200: Input the preprocessed slices into the first segmentation network to segment the aorta, and obtain the aortic region slice map and aortic category; Step S300: Input the aortic region slice image into the second segmentation network based on U-Net for intimal flap segmentation. Its encoder includes multiple feature extraction layers with progressive downsampling. Each feature extraction layer introduces a channel attention module SEBlock. The SEBlock is used to obtain the global response information of each channel and learn the weight of each channel by performing global pooling on the current feature map, so as to adaptively recalibrate the features of different channels. Its decoder includes multiple progressively upsampled feature recovery layers and fuses features from the corresponding layers of the encoder to obtain feature maps. The decoder introduces a spatial attention module CBAM, which generates a spatial attention heatmap by performing pooling and convolution operations on the feature map in the spatial dimension, enabling the network to focus on the subtle differences at the edge of the endometrial flap during the decoding stage. The decoder has deep supervision branches at different levels, which connect multiple intermediate feature layers of different resolutions and are used to apply auxiliary supervision constraints to the intermediate feature outputs. Step S400: Calculate the positive percentage of aortic dissection for each aortic category, and perform DeBakey classification of aortic dissection based on the extent of dissection involvement.
2. The aortic dissection prediction method based on U-Net dual attention mechanism according to claim 1, characterized in that, Step S200 further includes: sorting the CTA slices according to the order of their corresponding slice layers in the CTA image to be detected along the stacking direction to form a slice sequence; using a sliding window to filter the junctions of different segments of the aorta, dividing the slice sequence into slice subsequences of different aortic categories, classifying each slice subsequence into the corresponding segment according to the number and shape of the aortic regions detected in the slice, and removing slice images in each slice subsequence that do not belong to the aortic category.
3. The aortic dissection prediction method based on U-Net dual attention mechanism according to claim 1, characterized in that, The first segmentation network is based on the U-Net network, and residual blocks are constructed at each block to prevent gradient vanishing and expand the receptive field.
4. The aortic dissection prediction method based on U-Net dual attention mechanism according to claim 3, characterized in that, The deep supervision branch connects at least three intermediate feature layers with different resolutions. Each deep supervision branch includes a convolutional mapping layer and an upsampling layer, so that the output of the intermediate feature layers is mapped to a spatial scale consistent with the final segmentation result.
5. The aortic dissection prediction method based on U-Net dual attention mechanism according to claim 4, characterized in that, In step 300, during the training phase of the second segmentation network, a knowledge distillation method is used. A dual-domain feature distillation framework is constructed to address the characteristics of significant differences in fine-grained features and blurred boundaries of the endometrial flap structure in CTA images. This framework includes two distillation paths: semantic domain distillation and structural domain distillation. This allows the student network to simultaneously learn the high-level semantic representation and low-level structural details of the teacher model during training.
6. The aortic dissection prediction method based on U-Net dual attention mechanism according to claim 5, characterized in that, Step S300 also includes: for the channel features output by each convolutional layer in the second segmentation network decoder and encoder, firstly calculate the normalized weight response value during the training phase as an importance index of the channel contribution; then, prune the convolutional kernels corresponding to the convolutional channels whose importance scores are lower than a preset threshold and permanently remove them from the network structure so that the channel is no longer calculated during the inference phase.
7. The aortic dissection prediction method based on U-Net dual attention mechanism according to claim 6, characterized in that, The aortic category mentioned in step 400 includes three anatomical segments: the ascending aorta, the aortic arch, and the descending aorta. Based on the distribution of the intimal flap and false lumen in each segment, the segment involved in the dissection is determined, providing a basis for the DeBakey classification.
8. The aortic dissection prediction method based on U-Net dual attention mechanism according to claim 7, characterized in that, The method further includes: performing three-dimensional vascular reconstruction based on the aortic segmentation map to generate a three-dimensional model of the aorta, and further performing three-dimensional vascular reconstruction again based on the intimal flap segmentation results to further improve the three-dimensional structural model of aortic dissection.
9. A predictive system for aortic dissection based on U-Net dual attention mechanism, characterized in that, The system has a program module corresponding to the steps of the method described in any one of claims 1 to 8, and executes the steps in the aortic dissection prediction method based on the U-Net dual attention mechanism described above when it is run.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program configured to, when invoked by a processor, implement the steps of the aortic dissection prediction method based on the U-Net dual attention mechanism as described in any one of claims 1 to 8.