Artificial intelligence assisted acute aortic syndrome cta automatic diagnosis system

Abstract

The present application relates to the medical image diagnosis technical field, specifically to an artificial intelligence assisted acute aortic syndrome CTA automatic diagnosis system. It aims to solve the problems of small branch lesions being easily covered by the main stem, long diagnosis time, and lack of standardization of primary diagnosis and treatment in the existing acute aortic syndrome CTA diagnosis. The technical scheme points include: adopting the core architecture of "two-stage segmentation + slice-level classification", introducing the "branch priority" logic to avoid missing small branch lesions, integrating a visual interface to support multi-format import and mode selection, and ensuring model versatility through cross-institutional verification. The main use is to assist clinicians in quickly and accurately diagnosing acute aortic syndrome, provide standardized diagnosis and treatment support, and be suitable for medical institutions at all levels, especially primary care, to help improve diagnosis and treatment efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of medical image information processing technology, and in particular to an artificial intelligence-assisted CTA automatic diagnostic system for acute aortic syndrome. Background Technology

[0002] With the development of artificial intelligence in medical imaging and cardiovascular diagnosis and treatment technologies, CTA imaging data for acute aortic syndrome (AAS) has enormous potential. This type of data contains aortic anatomy and lesion characteristics, serving as a core basis for accurate diagnosis. The lesion details contained in this data (such as intimal tear and the extent of intramural hematoma) have crucial clinical value. Through systematic analysis, AAS subtypes (aortic dissection (AD), intramural hematoma (IMH), penetrating ulcer (PAU)) and lesion regions can be clearly identified, helping to develop targeted treatment plans and promoting the transformation of AAS diagnosis and treatment from "experience-dependent" to "data-driven," from "passive treatment" to "proactive prediction," and from "rough assessment" to "precise regional localization." Currently, CTA diagnosis of AAS is still mainly based on manual interpretation. The imaging data has not formed a standardized integration system and is mostly scattered in different image storage systems. AAS diagnosis relies on doctors' experience, and there is a significant gap in accuracy between junior doctors and senior experts. In addition, manually completing the labeling of the 17 aortic zones and subtype determination is time-consuming, which is difficult to meet the emergency needs of AAS, where "the mortality rate increases by 1%-2% for every hour of delay". Furthermore, there are problems such as missed lesions and inaccurate risk prediction due to incomplete information interpretation.

[0003] Chinese invention CN120339681A discloses a method for early warning of acute aortic syndrome, comprising: acquiring and preprocessing aortic CT images; generating CTA images using a cascaded generative adversarial network model; extracting key information features from the CTA images using the deep learning classification model DenseNet, and outputting aortic syndrome classification results. It can generate high-quality CTA images without the use of contrast agents, reducing the risks to patients who may have undergone contrast agent examinations, and is particularly suitable for patients with impaired renal function or contrast agent allergies. However, this invention has the following shortcomings: it does not mention a precise segmentation scheme for aortic minor branches (such as the renal artery and brachiocephalic artery), thus failing to address the problem of missed diagnoses due to the low voxel ratio of minor branches; it does not involve a visualization and interpretation mechanism for AI diagnostic results, resulting in a "black box" problem and poor clinical trust and interpretability; it does not explain multi-center data or cross-device verification, which may lead to single-center bias; it only outputs classification results without lesion localization, failing to provide more detailed evidence for treatment plans; and it lacks practical functions such as system integration (e.g., GUI interface, PACS integration), clinical decision rules (e.g., branch priority), or automatic report generation, failing to meet the actual needs of clinical practice.

[0004] Chinese invention CN120413004A provides a method and system for automatic measurement of aortic morphological features and clinical decision support based on artificial intelligence, applicable to three-dimensional medical image analysis of the aorta and its major branches. The method includes: image data acquisition and preprocessing, automatic vascular structure segmentation, centerline extraction and key point localization, automatic multi-parameter measurement, risk assessment and clinical decision support, data standardization output, and system feedback optimization. The corresponding system includes an image acquisition and processing module, a segmentation module, a centerline extraction module, a parameter measurement module, a decision support module, a data output module, and a validation and optimization module. Through deep learning models and morphological computation methods, it achieves efficient identification and quantitative analysis of vascular structures, and outputs individualized risk recommendations and surgical planning based on statistical and machine learning models. However, this invention only analyzes the "major branches" of the aorta, without mentioning a precise segmentation scheme for smaller branches (such as the renal artery and brachiocephalic artery), thus failing to address the issue of missed diagnoses due to the low voxel percentage of smaller branches; it lacks a visualization and interpretation mechanism for AI diagnostic results, presenting a "black box" problem with poor clinical trust and interpretability; it does not mention priority processing rules for lesions in smaller branches; it cannot prevent lesions in smaller branches from being masked by signals from the main trunk, thus failing to meet the need for precise identification of minute lesions in emergency treatment; it focuses on aortic morphology measurement and risk assessment, without explicitly mentioning the ability to accurately classify subtypes of acute aortic syndrome (AAS) (AD / IMH / PAU); it cannot achieve slice-level precise classification of the three subtypes, thus failing to address the diagnostic pain points of AAS; it is not optimized for the emergency need where "mortality increases by 1%-2% for every hour of delay in AAS," resulting in long diagnosis times that do not meet the timeliness requirements of emergency treatment; the system module description is too general, without explicitly mentioning practical functions such as GUI interface, PACS integration, and automatic report generation, failing to meet the actual clinical operational needs. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide an artificial intelligence-assisted CTA automatic diagnostic system for acute aortic syndrome, which integrates "automated anatomical segmentation, precise subtype classification and clinical decision support". It adopts a core technical architecture of "two-stage segmentation + slice-level classification" to overcome clinical difficulties, significantly shorten the diagnosis time, reduce the risk of missing small branch lesions, and achieve diagnostic accuracy close to that of senior experts. It provides support for AAS emergency treatment and standardized diagnosis and treatment in primary medical institutions, and helps to improve both the efficiency and quality of diagnosis and treatment.

[0006] This invention provides an artificial intelligence-assisted CTA automated diagnostic system for acute aortic syndrome, the construction process of which includes:

[0007] S1. Data preparation, including:

[0008] S11. Select standard-compliant data from different sources as training and validation sets respectively. The data are CTA images of AAS patients with AD, IMH, and PAU subtypes, covering the complete anatomical range from the aortic root to the iliac artery bifurcation.

[0009] S12. Preprocess the data of the training set and the validation set. The preprocessing includes patient information protection, data quality screening, and data format unification to adapt to the training requirements of the subsequent deep learning model.

[0010] S12. Perform initial segmentation and subtype diagnostic labeling of the data into 17 regions, and verify the labeling through quality control to ensure its reliability; the 17 regions include 10 regions of the main aorta and 7 regions of the aortic minor branches;

[0011] S2. Segmentation Model Training: A two-stage segmentation strategy is adopted for training. The first stage is to perform accurate segmentation of the 10 zones of the main aorta, and the second stage is to optimize the segmentation of the 7 zones of the aortic minor branches, ultimately achieving full anatomical coverage of 17 zones.

[0012] The precise segmentation of the 10 zones of the main aorta specifically includes:

[0013] SA1, Data Input: Extract the backbone region of the preprocessed training set and divide it into an internal training set and an internal validation set according to the proportion. Patient-level hierarchical structure is adopted to avoid cross-subset contamination of data from the same patient.

[0014] SA2, Model Training: Based on the nnUNet-V2 framework, the full-resolution 3D processing mode is selected, the architecture is 3DPlainConvUNet, the training device is NVIDIA A6000GPU, and the model is trained on the internal training set to obtain a 10-region segmentation model.

[0015] Furthermore, during model training, regional overlap evaluation metrics are periodically calculated on the internal validation set to monitor training effectiveness; model convergence criteria are set, and an early stopping mechanism is triggered to terminate training when the evaluation metrics on the internal validation set show no improvement for multiple consecutive rounds; this ensures that the model's segmentation accuracy of the target region meets preset requirements, thereby guaranteeing the accuracy of the target region's anatomical boundaries.

[0016] The optimization of the seven sub-branch splits specifically includes:

[0017] SB1, Unified Mask Pre-training: The labels including the 10 regions of the main aorta and the 7 regions of the minor branches of the aorta are merged into a single mask. The basic segmentation model is pre-trained with the unified mask to allow the basic segmentation model to initially learn the overall structural features of the aorta in the "main trunk-branch" structure. This ensures that the DSC value of the single mask meets the requirements and generates a pre-trained 17-region segmentation mask.

[0018] SB2, Structural Prior Retraining: The pre-trained 17-region segmentation mask is converted into a numerical format consistent with the original CTA image. High-density regions are selected as an additional input channel. The main aortic trunk and small branches are input into the model respectively. At this time, there are a total of 8 segmentation models including the main aortic trunk, brachiocephalic artery, left common carotid artery, left subclavian artery, celiac trunk, superior mesenteric artery and bilateral renal arteries. The model is retrained. During retraining, the focus is on optimizing the gradient update of the branch regions and guiding the model to pay attention to the low voxel features of small branches, resulting in an 8-region segmentation model.

[0019] SB3. Validate the 8-region segmentation model using the internal validation set to ensure that the accuracy of the 8-region segmentation model meets the requirements;

[0020] SB4. The segmentation models of the 10 regions of the main aorta and the 7 regions of the minor branches (excluding the main aorta) are fused together. For the voxel overlap areas, the principle of "minor branches first" is adopted, so that the segmentation accuracy of the 17 regions of the aorta finally meets the requirements.

[0021] S3. Training of a slice-level AAS subtype classification model based on ResNet-18, including:

[0022] S31, Slice generation: Based on the 17-region segmentation mask obtained in SB1, the training set is sliced ​​layer by layer using the NiBabel library of Python to remove the background and retain only the mask-covered area to generate a rectangular PNG image.

[0023] S32, Subtype Splitting: The rectangular PNG images are divided into three independent datasets according to the lesion type: AD, IMH, and PAU. Each dataset is split into slice-level training set and slice-level validation set according to the proportion.

[0024] S33. Subtype classification model training: The ResNet-18 classification model is trained step by step using the three independent datasets to obtain the slice-level AAS subtype classification model.

[0025] S34. Grad-CAM Attention Heatmap Integration: A Grad-CAM module is added to the last convolutional layer of the slice-level AAS subtype classification model to calculate the gradient weights of the lesion region, generate a heatmap, and overlay it with the original slice for output, providing clinical interpretability.

[0026] S4. AAS-DSS System Integration and Clinical Application Validation: The system functions are integrated with a focus on clinical practicality, and the validation set is used for value validation. The integration of system functions involves integrating the 17-region segmentation model and the slice-level AAS subtype classification model into a visual interface to obtain the AAS-DSS system, namely the artificial intelligence-assisted CTA automatic diagnosis system for acute aortic syndrome.

[0027] Furthermore, in SA2 and SB1, the model parameter settings include: epochs=1000, batchsize=2, optimizer is Adam, initial lr=0.001, loss function is DiceLoss+CrossEntropyLoss, Z-score normalization preprocessing is used, mean=0, standard deviation=1.

[0028] Furthermore, the early stopping mechanism of SA2 is as follows: the Dice similarity coefficient (DSC) is calculated on the internal validation set every 50 epochs, and early stopping is triggered when there is no improvement in DSC for 20 consecutive epochs.

[0029] Furthermore, the training of the subtype classification model for S33 specifically includes:

[0030] Model initialization: Load the ImageNet pre-trained ResNet-18 weights, freeze the first 8 layers, and fine-tune the last 10 layers after unfreezing to adapt to AAS lesion features;

[0031] Training parameter settings: The training device was an NVIDIA RTX 5060 Ti GPU, with 100 epochs and a batch size of 32; the optimizer was Adam, with an initial lr of 0.001, paired with the ReduceLROnPlateau scheduler: patience = 5 and factor = 0.5; the loss function was CrossEntropyLos, used to calculate the difference between the model's predicted subtype labels (AD / IMH / PAU) and the true labels, guiding the direction of model optimization; image preprocessing included adjusting sliced ​​images to a uniform size, converting images from pixel matrices to tensor formats that can be processed by deep learning frameworks, and standardizing image pixel values ​​using the mean and standard deviation of the ImageNet dataset to eliminate brightness / contrast differences between different images and improve model stability;

[0032] Early stopping mechanism setting: Training is stopped when the loss of the slice-level validation set does not decrease for 5 consecutive epochs to avoid overfitting and ensure that the final classification indicators of the three subtypes meet the requirements. The classification indicators include accuracy, false negative rate and clinical value of low FNR.

[0033] Furthermore, the AAS-DSS system integration in S4 specifically includes:

[0034] Interface Development: The GUI is built based on the Tkinter framework and consists of three parts: image loading area, parameter setting area, and result display area.

[0035] The image loading area supports importing DICOM / .nii.gz format and automatically recognizes the dimensions and layer thickness of CTA images;

[0036] The parameter setting area is used to select between fast mode and detailed mode, and by default loads the pre-trained weights of nnUNet-V2 and ResNet-18; wherein, the fast mode only outputs the diagnostic results, and the detailed mode outputs the segmentation mask + heatmap;

[0037] The results display area displays 17-zone color segmentation maps, subtype diagnostic scores, Grad-CAM heatmaps, and branch priority prompts in real time.

[0038] The logic of branch priority is as follows: in the overlapping area of ​​Label 2 and Label 1, the system prioritizes the diagnostic results of the branch to avoid the lesions of small branches being covered by the main trunk.

[0039] Furthermore, the clinical application validation in S4 specifically includes:

[0040] Comparative verification: The verification set was selected, and the AAS-DSS system, junior doctors, and senior doctors respectively completed the diagnosis of 17 zones, and the diagnosis time and accuracy were recorded;

[0041] Output results: The AAS-DSS system automatically generates a diagnostic report, which includes a 17-zone lesion distribution table, key lesion slices annotated with hot icons, examination recommendations, and assists clinicians in developing surgical or interventional plans.

[0042] The technical solution provided by this invention has at least the following technical effects:

[0043] 1. Adopting a core architecture of "two-stage segmentation + slice-level classification", the diagnosis time is significantly shortened, the risk of missing small branch lesions is reduced, and the diagnostic accuracy is close to that of senior experts;

[0044] 2. Introduce a "branch-first" logic to prevent lesions in small branches from being obscured by the main trunk, thereby improving the comprehensiveness and accuracy of lesion detection;

[0045] 3. The integrated visual interface supports importing multiple formats and selecting modes, making it easy to operate, meeting the needs of different clinical scenarios, and assisting in rapid decision-making;

[0046] 4. Ensure the model's universality through cross-institutional validation, and help primary healthcare institutions achieve standardized diagnosis and treatment;

[0047] 5. Combining Grad-CAM heatmaps provides clinical interpretability and enhances physicians' confidence in diagnostic results.

[0048] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0049] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0050] Figure 1 This is a schematic diagram of the construction process of the system of the present invention;

[0051] Figure 2 This is a table showing the correspondence between aortic zonal branches defined by the 17-zone label in this embodiment of the invention.

[0052] Figure 3 This is a schematic diagram illustrating the training process of the segmentation model of the AAS-DSS system according to an embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram illustrating the training process of the classification model of the AAS-DSS system according to an embodiment of the present invention;

[0054] Figure 5 This is a typical case of type A aortic dissection involving branch vessels, as described in this embodiment of the invention: AAS-DSS slice diagnosis and Grad-CAM attention heatmap. Detailed Implementation

[0055] This application provides an AI-assisted CTA automated diagnostic system for acute aortic syndrome, integrating "automated anatomical segmentation, precise subtype classification, and clinical decision support." It adopts a core technical architecture of "two-stage segmentation + slice-level classification" to overcome clinical difficulties, significantly shortening diagnosis time, reducing the risk of missing small branch lesions, and achieving diagnostic accuracy close to that of senior experts. This system provides support for AAS emergency treatment and standardized diagnosis and treatment in primary healthcare institutions, helping to improve both the efficiency and quality of diagnosis and treatment.

[0056] The specific ideas behind this invention are as follows:

[0057] First, a large sample of high-quality data was used to lay the foundation for the model's reliability. CTA images (including AD, IMH, and PAU subtypes) of 586 AAS patients meeting the criteria from the first data source were selected as the training set. Patient information was strictly anonymized in accordance with ethical guidelines to ensure data coverage of the complete anatomical range from the aortic root to the iliac artery bifurcation. Simultaneously, data from 51 independent external centers (using different CT scanners) from the second data source were included for validation to avoid single-center data bias and ensure the model's cross-institutional applicability.

[0058] Secondly, a core technical architecture of "two-stage segmentation + slice-level classification" is adopted to overcome clinical challenges. In the segmentation stage, the nnUNet-V2 framework is used for phased processing: first, accurate segmentation of the 10 zones (SVS / STSZones0-9) of the aortic trunk is achieved; then, the segmentation of 7 small branches (such as the brachiocephalic artery and renal artery) is optimized through the strategy of "unified mask pre-training + structural prior re-training" to solve the problem of small vessel omission caused by voxel imbalance, and finally achieve full anatomical coverage of 17 zones; in the classification stage, more than 330,000 slice-level images are extracted, and ResNet-18 is used to train AD, IMH, and PAU subtype classification models respectively. Grad-CAM attention heatmap is used to visualize key lesion areas, balancing diagnostic accuracy and clinical interpretability.

[0059] Finally, the system's functions were integrated and its value validated with a focus on clinical practicality. The segmentation and classification modules were integrated into the TkinterGUI interface, achieving full automation of the "one-click loading - automatic processing - result output" process. At the same time, a "branch priority" rule was set to ensure that lesions in small branches are not masked by the main trunk signal. Through comparative validation with senior (10 years) and junior (3 years) radiologists, the diagnostic accuracy was reduced by 80% while maintaining a level close to that of senior experts, significantly improving efficiency. Ultimately, a standardized AAS diagnostic tool was provided for grassroots and resource-scarce areas, helping to improve both the timeliness and quality of emergency treatment. Example

[0060] like Figure 1 As shown, this embodiment provides an artificial intelligence-assisted CTA automatic diagnostic system for acute aortic syndrome, the construction process of which includes:

[0061] The acquisition, preprocessing, and annotation of S1 and AAS-CTA image data provide a high-quality, standardized data foundation for subsequent model training, addressing the poor generalization issues caused by incomplete data coverage and inconsistent annotations in existing technologies. This includes:

[0062] S11. Select standard-compliant data from different sources as the training set and validation set, respectively. The data consists of CTA images of AAS patients including AD, IMH, and PAU subtypes, covering the complete anatomical range from the aortic root to the iliac artery bifurcation; wherein...

[0063] Training set: CTA images of 586 AAS patients collected from Fuzhou University Affiliated Provincial Hospital between June 2015 and June 2024 were collected. All images were collected using a Philips Brilliance 64-slice CT scanner. The parameters were uniform: detector collimation 64×0.625mm, pitch 0.90, rotation time 0.60s, tube voltage 120kVp, automatic tube current (target 300mAs), reconstruction slice thickness 0.625mm, and increment 0.3mm. The images were designed to cover the complete anatomical range of "aortic root (Warshall sinus) - iliac artery bifurcation" and include three subtypes: AD (112 cases), IMH (86 cases), and PAU (113 cases).

[0064] Validation set: 51 independent CTA images were collected from the Second Affiliated Hospital of Zhengzhou University from June 2020 to December 2022 using a Siemens 256-slice dual-source CT scanner (simulating cross-institutional equipment differences). The case composition was 38 cases of AD, 10 cases of IMH, and 12 cases of PAU, which met the inclusion and exclusion criteria consistent with the training set (inclusion: ≥18 years old, complete images; exclusion: <18 years old, congenital aortic malformation, lack of informed consent).

[0065] S12. Preprocess the data of the training set and the validation set. The preprocessing includes patient information protection, data quality screening, and data format unification to adapt to the training requirements of the subsequent deep learning model.

[0066] Patient information protection: The Pandas library of Python is used to anonymize patient information such as name and ID number, and only the image ID and lesion label are retained, which complies with the Declaration of Helsinki and the ethics approval of Fujian Provincial Hospital (K2025-02-111).

[0067] Quality screening: Two junior radiologists (3 years of experience) independently removed images with severe artifacts and incomplete anatomical range. In case of disagreement, a senior radiologist (10 years of experience) made the judgment. Finally, 586+51 valid data were retained.

[0068] Format conversion: Convert DICOM format images to .nii.gz format, and use the SimpleITK library to adjust the pixel value range (-1000~400HU) to avoid CT value differences affecting model training.

[0069] S12. Perform initial segmentation and subtype diagnostic labeling of the data into 17 regions, and verify the labeling through quality control to ensure its reliability; the 17 regions include 10 regions of the main aorta and 7 regions of the aortic minor branches;

[0070] Annotation tool: 3D Slicer 5.8.1 open-source platform, based on the extended SVS / STS classification system (10 main branches + 7 sub-branches) to define 17 area labels, such as... Figure 2 As shown, this is the correspondence table for Labels 1-17.

[0071] Labeling process: ① Senior physicians (10 years of experience) complete the initial segmentation and subtype diagnosis (AD / IMH / PAU) labeling of 17 regions; ② Two junior physicians blind review the labeling results and calculate the Kappa coefficient (≥0.85 required); ③ Cases with Kappa <0.85 are jointly adjudicated by 3 physicians, and the final result is used as the "gold standard" for model training.

[0072] Quality control verification: 20% of the labeled data are randomly selected and reviewed by another senior doctor from another hospital (10 years of experience). The accuracy rate must be ≥98% to ensure the reliability of the labeling.

[0073] S2, Segmentation Model Training: (e.g.) Figure 3 As shown, a two-stage segmentation strategy was used for training. The first stage was to perform precise segmentation of the 10 regions of the aortic trunk, and the second stage was to optimize the segmentation of the 7 regions of the aortic minor branches, ultimately achieving full anatomical coverage of 17 regions.

[0074] The precise segmentation of the 10 zones of the main aorta specifically includes:

[0075] SA1, Data Input: Extract the backbone region (Label 1, 3, 5, 7, 8, 9, 10, 12, 14, 17, corresponding to SVS / STS Zones 0-9) of the preprocessed training set (586 images from the first data source) and split it into an internal training set and an internal validation set. Patient-level stratification is used to avoid cross-subset contamination of data from the same patient.

[0076] SA2, Model Training: Based on the nnUNet-V2 framework, the full-resolution 3D processing mode is selected, the architecture is 3DPlainConvUNet, the training device is NVIDIA A6000GPU, and the model is trained on the internal training set to obtain a 10-region segmentation model.

[0077] The model parameters are set as follows: epochs=1000, batchsize=2, optimizer is Adam, initial lr=0.001, loss function is DiceLoss+CrossEntropyLoss, Z-score normalization preprocessing is used, mean=0, standard deviation=1.

[0078] Furthermore, during model training, regional overlap evaluation metrics are periodically calculated on the internal validation set to monitor training effectiveness; model convergence criteria are set, and an early stopping mechanism is triggered to terminate training when the evaluation metrics on the internal validation set show no improvement for multiple consecutive rounds; this ensures that the model's segmentation accuracy of the target region meets preset requirements, thereby guaranteeing the accuracy of the target region's anatomical boundaries.

[0079] The early stopping mechanism of SA2 is as follows: the Dice similarity coefficient (DSC) is calculated on the internal validation set every 50 epochs, and early stopping is triggered when there is no improvement in DSC for 20 consecutive epochs. In a specific embodiment, the final average DSC of the 10 regions of the backbone reaches 0.870 (Label 1 / Zone 0: 0.937, Label 17 / Zone 9: 0.929), ensuring accurate backbone anatomical boundaries.

[0080] The second stage of segmentation optimization of the 7 small branches is used to address the core pain points of existing technologies: the voxel ratio of the small branches (Label2 / infrared artery, Label4 / left common carotid artery, etc.) is only 15%-20% of the aortic trunk, and single-stage training is easily masked by the trunk signal. The DSC of the branches in existing technologies is mostly <0.85.

[0081] The optimization of the seven sub-branch splits specifically includes:

[0082] SB1, Unified Mask Pre-training: The labels including the 10 regions of the aortic trunk and the 7 regions of the aortic minor branches are merged into a single mask. The basic segmentation model is pre-trained with the unified mask to allow the basic segmentation model to initially learn the overall structural features of the aorta in the "trunk-branch" pattern. The DSC value of the single mask meets the requirements (in a specific embodiment, DSC=0.959), generating a pre-trained 17-region segmentation mask.

[0083] The model parameters were set to be the same as those in the first stage for training the basic segmentation model, including: epochs=1000, batchsize=2, optimizer is Adam, initial lr=0.001, loss function is DiceLoss+CrossEntropyLoss, Z-score normalization preprocessing is used, mean=0, standard deviation=1.

[0084] SB2, Structural Prior Retraining: The pre-trained 17-region segmentation mask is converted into a numerical format consistent with the original CTA image. High-density regions are selected as an additional input channel. The main aortic trunk and small branches are input into the model respectively. At this time, there are a total of 8 segmentation models including the main aortic trunk, brachiocephalic artery, left common carotid artery, left subclavian artery, celiac trunk, superior mesenteric artery and bilateral renal arteries. The model is retrained. During retraining, the focus is on optimizing the gradient update of the branch regions and guiding the model to pay attention to the low voxel features of small branches, resulting in an 8-region segmentation model.

[0085] SB3. Validate the 8-region segmentation model using the internal validation set to ensure that the accuracy of the 8-region segmentation model meets the requirements;

[0086] SB4. The segmentation models of the 10 regions of the main aorta and the 7 regions of the minor branches (excluding the main aorta) are fused together. For the voxel overlap areas, the principle of "minor branches first" is adopted, so that the segmentation accuracy of the 17 regions of the aorta finally meets the requirements.

[0087] In one specific embodiment, the DSC of the small branches in the internal validation set was improved to 0.889-0.937 (Label2: 0.926, Label13 / superior mesenteric artery: 0.937, Label15 / left renal artery: 0.898), which solved the voxel imbalance problem. Finally, the average DSC of all 17 regions reached 0.891.

[0088] S3, such as Figure 4 As shown, the training of the slice-level AAS subtype classification model based on ResNet-18 includes:

[0089] S31. Slice generation: Based on the 17-region segmentation mask obtained in SB1, the training set is sliced ​​layer by layer using the NiBabel library of Python. The background is removed and only the mask-covered area is retained to generate rectangular PNG images. Specifically, 339,798 rectangular PNG images of 224×224 pixels can be generated.

[0090] S32. Subtype Splitting: The rectangular PNG images are divided into three independent datasets according to lesion type: AD, IMH, and PAU. Each dataset is further split into slice-level training and slice-level validation sets according to a specific ratio. The three independent datasets, AD, IMH, and PAU, are as follows:

[0091] AD / Non-AD (48,683 / 291,115 images), IMH / Non-IMH (12,453 / 327,345 images), PAU / Non-PAU (14,531 / 325,267 images).

[0092] S33. Subtype Classification Model Training: The ResNet-18 classification model is trained step-by-step using the three independent datasets to obtain a slice-level AAS subtype classification model; specifically including:

[0093] Model initialization: Load the ImageNet pre-trained ResNet-18 weights, freeze the first 8 layers, and fine-tune the last 10 layers after unfreezing to adapt to AAS lesion features;

[0094] Training parameter settings: The training device is an NVIDIA RTX 5060 Ti GPU, epochs=100, batch size=32; the optimizer is Adam, with an initial lr=0.001, paired with the ReduceLROnPlateau scheduler: patience=5, factor=0.5; the loss function is CrossEntropyLos, used to calculate the difference between the model's predicted subtype labels (AD / IMH / PAU) and the true labels, guiding the direction of model optimization; image preprocessing is "Resize→Tensor transformation→ImageNet normalization (mean [0.485,0.456,0.406], standard deviation [0.229,0.224,0.225])", which includes adjusting sliced ​​images to a uniform size, converting images from pixel matrices to tensor formats that can be processed by deep learning frameworks, and using the mean and standard deviation of the ImageNet dataset to standardize image pixel values, eliminating brightness / contrast differences between different images, and improving model stability;

[0095] Early stopping mechanism setting: Training is stopped when the loss of the slice-level validation set does not decrease for 5 consecutive epochs to avoid overfitting and ensure that the final classification indicators of the three subtypes meet the requirements. The classification indicators include accuracy, false negative rate and clinical value of low FNR.

[0096] In a specific implementation, the final three subtype classification indicators are as follows: AD (Accuracy=0.976, FNR=1.0%), IMH (Accuracy=0.990, FNR=0.41%), and PAU (Accuracy=0.972, FNR=1.77%). A low FNR ensures that acute cases are not missed.

[0097] S34, Grad-CAM Attention Heatmap Integration: Its technical purpose is to solve the AI ​​"black box" problem, allowing doctors to intuitively see the key evidence of model diagnosis, which meets the clinical interpretability requirements;

[0098] Implementation: A Grad-CAM module is added to the last convolutional layer of the slice-level AAS subtype classification model to calculate the gradient weights of the lesion region, generate a heatmap (red areas represent high-weight lesion features, such as the intimal flap in AD and the crescent-shaped high-density area in IMH), and overlay it with the original slice for output, providing clinical interpretability. See example. Figure 5 As shown, the thermal image of the Type A interlayer focuses on the tear in the inner membrane, with a weight value > 0.8.

[0099] S4. AAS-DSS System Integration and Clinical Application Validation: The system functions are integrated with a focus on clinical practicality, and the validation set is used for value validation. The integration of system functions involves integrating the 17-region segmentation model and the slice-level AAS subtype classification model into a visual interface to obtain the AAS-DSS system, namely the artificial intelligence-assisted CTA automatic diagnosis system for acute aortic syndrome.

[0100] This step integrates the segmentation and classification modules into a visual interface to achieve "one-click automated diagnosis," and verifies the system's practicality through cross-institutional validation. The specific steps are as follows:

[0101] The AAS-DSS system integration specifically includes:

[0102] Interface Development: The GUI is built based on the Tkinter framework and consists of three parts: "Image Loading Area," "Parameter Setting Area," and "Result Display Area."

[0103] The image loading area supports importing DICOM / .nii.gz format images and automatically recognizes CTA image dimensions and layer thickness.

[0104] The parameter settings area allows you to select "Fast Mode (outputs diagnostic results only)" or "Detailed Mode (outputs segmentation mask + heatmap)". By default, it loads pre-trained nnUNet-V2 (segmentation) and ResNet-18 (classification) weights.

[0105] The results display area shows in real time the 17-zone color segmentation map, subtype diagnostic score (out of 100), Grad-CAM heatmap, and "branch priority" prompt;

[0106] The core rule is to implement the "branch priority" logic: in overlapping areas such as Label 2 (innocent artery) and Label 1 (Zone 0), the system prioritizes the diagnostic results of the branch (if the branch is marked as IMH positive, the corresponding area of ​​the main trunk will not be marked as negative), so as to avoid small branch lesions being covered by the main trunk.

[0107] The clinical application validation specifically includes:

[0108] Comparative verification: The verification set was selected, and the AAS-DSS system, junior doctors, and senior doctors respectively completed the diagnosis of 17 zones, and the diagnosis time and accuracy were recorded;

[0109] Output results: The AAS-DSS system automatically generates a diagnostic report, which includes a 17-zone lesion distribution table, key lesion slices marked with hot icons, and examination suggestions (such as vascular ultrasound to confirm branch blood flow) to assist in the clinical development of surgical or interventional plans.

[0110] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An AI-assisted automated CTA diagnostic system for acute aortic syndrome, characterized in that: The construction process includes: S1. Data preparation, including: S11. Select standard-compliant data from different sources as training and validation sets respectively. The data are CTA images of AAS patients with AD, IMH, and PAU subtypes, covering the complete anatomical range from the aortic root to the iliac artery bifurcation. S12. Preprocess the data of the training set and the validation set. The preprocessing includes patient information protection, data quality screening, and data format unification to adapt to the training requirements of the subsequent deep learning model. S13. Perform initial segmentation and subtype diagnostic annotation of the data into 17 regions, and verify the annotation through quality control to ensure its reliability; the 17 regions include 10 regions of the main aorta and 7 regions of the aortic minor branches; S2. Segmentation Model Training: A two-stage segmentation strategy is adopted for training. The first stage is to accurately segment the 10 regions of the main aorta, and the second stage is to optimize the segmentation of the 7 regions of the aortic minor branches, ultimately achieving full anatomical coverage of 17 regions. When fusing the segmentation of the 10 regions of the main aorta and the 7 regions of the minor branches into a model, the principle of "minor branches first" is adopted for the voxel overlap areas. S3. Training of slice-level AAS subtype classification model based on ResNet-18; S4. AAS-DSS System Integration and Clinical Application Validation: The system functions are integrated with a focus on clinical practicality, and the validation set is used for value validation. The integration of system functions involves integrating the 17-region segmentation model and the slice-level AAS subtype classification model into a visual interface to obtain the AAS-DSS system, namely the artificial intelligence-assisted CTA automatic diagnosis system for acute aortic syndrome.

2. The system according to claim 1, characterized in that: The precise segmentation of the 10 zones of the main aorta specifically includes: SA1, Data Input: Extract the backbone region of the preprocessed training set and divide it into an internal training set and an internal validation set according to the proportion. Patient-level hierarchical structure is adopted to avoid cross-subset contamination of data from the same patient. SA2, Model Training: Based on the nnUNet-V2 framework, the full-resolution 3D processing mode is selected, the architecture is 3DPlainConvUNet, the training device is NVIDIA A6000GPU, and the model is trained on the internal training set to obtain a 10-region segmentation model. Furthermore, during model training, regional overlap evaluation metrics are periodically calculated on the internal validation set to monitor training effectiveness; model convergence criteria are set, and an early stopping mechanism is triggered to terminate training when the evaluation metrics on the internal validation set show no improvement for multiple consecutive rounds; this ensures that the model's segmentation accuracy of the target region meets preset requirements, thereby guaranteeing the accuracy of the target region's anatomical boundaries. The optimization of the seven sub-branch splits specifically includes: SB1, Unified Mask Pre-training: The labels including the 10 regions of the aortic trunk and the 7 regions of the aortic minor branches are merged into a single mask. The basic segmentation model is pre-trained with the unified mask to allow the basic segmentation model to initially learn the overall structural features of the aorta in the "trunk-branch" pattern. This ensures that the DSC value of the single mask meets the requirements and generates a pre-trained 17-region segmentation mask. SB2, Structural Prior Retraining: The pre-trained 17-region segmentation mask is converted into a numerical format consistent with the original CTA image. High-density regions are selected as an additional input channel. The main aortic trunk and small branches are input into the model respectively. At this time, there are a total of 8 segmentation models including the main aortic trunk, brachiocephalic artery, left common carotid artery, left subclavian artery, celiac trunk, superior mesenteric artery and bilateral renal arteries. The model is retrained. During retraining, the focus is on optimizing the gradient update of the branch regions and guiding the model to pay attention to the low voxel features of small branches, resulting in an 8-region segmentation model. SB3. Validate the 8-region segmentation model using the internal validation set to ensure that the accuracy of the 8-region segmentation model meets the requirements; SB4. The segmentation models of the 10 regions of the main aorta and the 7 regions of the minor branches are fused together. For the voxel overlap areas, the principle of "minor branches first" is adopted to finally achieve the required segmentation accuracy of the 17 regions of the aorta. The training of the ResNet-18-based slice-level AAS subtype classification model specifically includes: S31, Slice generation: Based on the 17-region segmentation mask obtained in SB1, the training set is sliced ​​layer by layer using the NiBabel library of Python to remove the background and retain only the mask-covered area to generate a rectangular PNG image. S32, Subtype Splitting: The rectangular PNG images are divided into three independent datasets according to the lesion type: AD, IMH, and PAU. Each dataset is split into slice-level training set and slice-level validation set according to the proportion. S33. Subtype classification model training: The ResNet-18 classification model is trained step by step using the three independent datasets to obtain the slice-level AAS subtype classification model. S34. Grad-CAM Attention Heatmap Integration: A Grad-CAM module is added to the last convolutional layer of the slice-level AAS subtype classification model to calculate the gradient weights of the lesion region, generate a heatmap, and overlay it with the original slice for output, providing clinical interpretability.

3. The system according to claim 2, characterized in that: In SA2 and SB1, the model parameters are set as follows: epochs=1000, batchsize=2, optimizer is Adam, initial lr=0.001, loss function is DiceLoss+CrossEntropyLoss, Z-score normalization preprocessing is used, mean=0, standard deviation=1.

4. The system according to claim 2, characterized in that: The early stopping mechanism of SA2 is as follows: the Dice similarity coefficient (DSC) is calculated on the internal validation set every 50 epochs, and early stopping is triggered when there is no improvement in DSC for 20 consecutive epochs.

5. The system according to claim 2, characterized in that: The training of the subtype classification model for S33 specifically includes: Model initialization: Load the ImageNet pre-trained ResNet-18 weights, freeze the first 8 layers, and fine-tune the last 10 layers after unfreezing to adapt to AAS lesion features; Training parameter settings: The training device was an NVIDIA RTX 5060 Ti GPU, with 100 epochs and a batch size of 32; the optimizer was Adam, with an initial lr of 0.001, paired with the ReduceLROnPlateau scheduler: patience = 5 and factor = 0.5; the loss function was CrossEntropyLos, used to calculate the difference between the model's predicted subtype labels (AD / IMH / PAU) and the true labels, guiding the direction of model optimization; image preprocessing included adjusting sliced ​​images to a uniform size, converting images from pixel matrices to tensor formats that can be processed by deep learning frameworks, and standardizing image pixel values ​​using the mean and standard deviation of the ImageNet dataset to eliminate brightness / contrast differences between different images and improve model stability; Early stopping mechanism setting: Training is stopped when the loss of the slice-level validation set does not decrease for 5 consecutive epochs to avoid overfitting and ensure that the final classification indicators of the three subtypes meet the requirements. The classification indicators include accuracy, false negative rate and clinical value of low FNR.

6. The system according to claim 1, characterized in that: The AAS-DSS system integration in S4 specifically includes: Interface Development: The GUI is built based on the Tkinter framework and consists of three parts: image loading area, parameter setting area, and result display area. The image loading area supports importing DICOM / .nii.gz format and automatically recognizes the dimensions and layer thickness of CTA images; The parameter setting area is used to select between fast mode and detailed mode, and by default loads the pre-trained weights of nnUNet-V2 and ResNet-18; wherein, the fast mode only outputs the diagnostic results, and the detailed mode outputs the segmentation mask + heatmap; The results display area displays 17-zone color segmentation maps, subtype diagnostic scores, Grad-CAM heatmaps, and branch priority prompts in real time. The logic of branch priority is as follows: in the overlapping area of ​​Label 2 and Label 1, the system prioritizes the diagnostic results of the branch to avoid the lesions of small branches being covered by the main trunk.

7. The system according to claim 1, characterized in that: The clinical application validation in S4 specifically includes: Comparative verification: The verification set was selected, and the AAS-DSS system, junior doctors, and senior doctors respectively completed the diagnosis of 17 zones, and the diagnosis time and accuracy were recorded; Output results: The AAS-DSS system automatically generates a diagnostic report, which includes a 17-zone lesion distribution table, key lesion slices annotated with hot icons, examination recommendations, and assists clinicians in developing surgical or interventional plans.

Images