Development and validation of an AI system for AAS CTA diagnosis and mapping

Figures

• 

  Figure 1. AAS-DSS diagnostic process The schematic depicts the end-to-end clinical workflow for AAS management enhanced by the AAS-DSS, structured into three sequential phases. (A) Initial clinical evaluation and imaging acquisition: patients with acute pain symptoms present to the hospital, undergo clinical assessment, and receive vascular CTA as the standard imaging modality for suspected AAS. (B) Automated AI analysis using the AAS-DSS: (1) input: raw aortic CTA imaging data are imported into the system; (2) processing: the AAS-DSS platform performs automated analysis of the imaging data; (3) segmentation: the system generates high-precision 3D segmentation of the aorta to isolate the vessel structure; and (4) diagnosis: based on the segmented aorta, the system outputs quantitative and visual diagnostic results, including interpretive heatmaps and anatomical assessments. (C) Clinical review and treatment planning: the AAS-DSS generates a structured diagnostic report, which is reviewed by clinicians to interpret findings, validate diagnoses, and formulate evidence-based, personalized treatment strategies for the patient.

• 

  Figure 2. Training of AAS-DSS aortic segmentation network The schematic illustrates the three-stage training strategy for the multi-zone aortic segmentation model. Stage I: Preprocessing; raw CTA images are processed to generate a dilated aortic mask, followed by targeted cropping to retain the entire aortic course from the aortic root to the iliac bifurcation, reducing background noise and ensuring the model’s focus on the target region. Stage II: Main trunk segmentation; preprocessed CTA images are paired with manually annotated 10-zone aortic trunk labels (legend, bottom right) and trained using nnU-Net to generate the segmented main aortic trunk. Stage III: Refined small-branch segmentation; a two-step strategy addresses the class imbalance between the large aortic trunk and small branches: (1) initial training uses a unified label (combining 10 trunk zones and 7 branch zones) to capture the full aortic region; (2) the initial full-region segmentation result (encoded at CT = 1,000) is co-input with the original CTA image, paired with “integral trunk + independent branch labels” for fine-tuning, improving segmentation accuracy of small branches.

• 

  Figure 3. Training of AAS-DSS aortic classification network The schematic shows the dataset composition and model interpretability for the three-pathology classification task. Left: schematic of the 17-zone aortic anatomy, defining the regions for targeted classification. Middle: representative cross-sectional CTA images and corresponding labels for each class: acute aortic dissection (AD, n = 48,683 slices), intramural hematoma (IMH, n = 12,453 slices), penetrating atherosclerotic ulcer (PAU, n = 14,531 slices), and their matched non-pathological counterparts. Right: Grad-CAM attention heatmaps derived from the fine-tuned ResNet18 model, highlighting the critical image regions driving classification decisions for AD, IMH, and PAU, demonstrating the model’s focus on pathological features.

• 

  Figure 4. Diagnostic performance evaluation of AAS-DSS for acute aortic syndrome subtypes on internal validation set (A–F) Classification performance for each acute aortic syndrome subtype. (A) Confusion matrix for AD classification, illustrating the model’s ability to correctly identify true-positive and true-negative cases. (B) Receiver operating characteristic (ROC) curve for AD classification, with area under the curve (AUC) quantifying overall discriminative performance. (C) Confusion matrix for IMH classification. (D) ROC curve for IMH classification. (E) Confusion matrix for PAU classification. (F) ROC curve for PAU classification. Abbreviations: AD, aortic dissection; IMH, intramural hematoma; PAU, penetrating aortic ulcer.

• 

  Figure 5. Zone-specific diagnostic performance of the AAS-DSS across 17 aortic segments in the external validation cohort The figure presents subtype-specific performance of the AAS-DSS across 17 standardized aortic zones defined by the Society for Vascular Surgery/Society of Thoracic Surgeons (SVS/STS). Labels correspond to the following segments: (1) zone 0 (aortic root to innominate artery); (2) innominate artery; (3) zone 1 (innominate artery to left common carotid artery); (4) left common carotid artery; (5) zone 2 (left common carotid artery to left subclavian artery); (6) left subclavian artery; (7) zone 3 (left subclavian artery to 2 cm distal); (8) zone 4 (to T6 vertebral body); (9) zone 5 (to celiac artery); (10) zone 6 (celiac artery to superior mesenteric artery); (11) celiac artery; (12) zone 7 (superior mesenteric artery to proximal renal artery); (13) superior mesenteric artery; (14) zone 8 (renal artery segment); (15) left renal artery; (16) right renal artery; (17) zone 9 (distal renal artery to aortic bifurcation). Performance metrics (e.g., accuracy, sensitivity, and specificity) are stratified by segment to evaluate the model’s diagnostic capability across the entire…