Automated AI detection of thoracic aortic dissection on CT imaging

Tobias Norajitra, Michael Baumgartner, Lucas R. Cusumano, Jesus G. Ulloa, Caterina Rizzo, Florian Haag, Alexander Hertel, Nils Rathmann, Steffen J. Diehl, Stefan O. Schoenberg, Klaus H. Maier-Hein, Johann S. RinkEuropean 2025 · DOI

TL;DR

Researchers have developed a robust AI pipeline using the nnU-Net framework to automatically detect and sub-classify thoracic aortic dissections (AD) from CT scans. The model achieved over 97% AUROC on both internal and external datasets, successfully identifying 93.3% of "unsuspected" cases that were not originally flagged for AD before imaging.

What Problem Does This Paper Address?

Acute aortic dissection (AD)—a life-threatening tear in the inner layer of the aorta—requires immediate diagnosis to prevent fatal complications like organ ischemia or aortic rupture. While Computed Tomography Angiography (CTA) is the gold standard for diagnosis, several challenges persist:

  1. Human Error: Nonspecific symptoms often lead to "unsuspected" cases where the radiologist isn't looking for AD, leading to potential misinterpretation or delay.
  2. Morphological Variety: AD presents with highly variable anatomy (Type A vs. Type B) and subtle features like thin dissection membranes.
  3. Data Heterogeneity: Imaging protocols and hardware vary wildly across hospitals, making it difficult for standard AI models to generalize.

The authors aimed to create an automated, "streamlined" AI tool that works reliably across multicenter data and different contrast phases.

Methods

The team opted for a semantic segmentation approach rather than simple image classification. By training the model to identify specific structures, the AI provides a "reason" for its detection, which is more useful for clinical validation.

  • Architecture: They utilized the nnU-Net framework (V1), a self-configuring method for deep learning-based biomedical image segmentation.
  • Labeling Strategy: The researchers tested three labeling configurations. The most successful (Configuration B) grouped 10 different anatomical features into five categories, prioritizing the "dissection membrane" to handle the class imbalance caused by its thinness.
  • Decision Logic: To convert segmentation into a binary "AD vs. non-AD" detection, they used a volume-based threshold. If at least two out of three key regions (ascending false lumen, descending false lumen, or membrane) exceeded a specific volume, the case was flagged as AD.
  • Stanford Classification: The model also attempted to distinguish between Type A (originating in the ascending aorta) and Type B dissections.

Datasets

The study utilized a mix of internal and public datasets to ensure generalizability:

  • Internal (Mannheim University Medical Centre): 2010–2023 data. Training set included 70 AD and 87 non-AD cases. The test set consisted of 106 cases, plus a separate "Atypical" test set of 16 cases with subtle anatomy.
  • External (Public): 100 AD cases from the ImageTBAD dataset (China) and 38 non-AD cases from the AVT dataset (multi-source).
  • Diversity: The data included various CT protocols (CTA, pulmonary artery contrast, mixed phases) and hardware from multiple vendors (Siemens, Philips).

Key Insights and Ideas

  • Catching the "Unsuspected": Perhaps the most significant finding was that the algorithm detected 14 out of 15 cases where AD was not suspected by the referring physician. This highlights the tool's potential as a "safety net" in busy emergency departments.
  • Segmentation as Detection: By formulating detection as a segmentation task, the model avoids the "black box" problem of standard classifiers. Clinicians can see exactly where the AI thinks the dissection is.
  • Prioritizing the Membrane: The authors found that prioritizing the thin dissection membrane during training (preventing it from being "overwritten" by larger lumen labels) was crucial for accuracy.
  • Robustness to Contrast: Unlike many previous models that require specific CTA timing, this pipeline remained accurate across different contrast phases, which is vital for real-world emergency imaging.

Key Results

The model demonstrated high performance across the board:

  • Internal Test Set: AUROC of 98.7% and AUPRC of 98.9%.
  • External Test Set: AUROC of 97.0% and AUPRC of 99.06%.
  • Sensitivity/Specificity: On external data, the model achieved 92.0% sensitivity and 100.0% specificity.
  • Human Comparison: The AI outperformed two advanced radiology residents in sensitivity on the internal test set (97.4% vs. 92.1% and 89.5%).
  • Speed: The average computation time was approximately 6 minutes and 38 seconds per case, making it feasible for near-real-time clinical support.

Limitations and Future Work

  • Type B Misclassification: The model occasionally misclassified Type B dissections as Type A if the intimal flap extended into the aortic arch—a known challenge even for human readers.
  • Retrospective Nature: The study was conducted on historical data. A prospective clinical trial is needed to see how the tool performs in a live "high-pressure" ER environment.
  • Dataset Size: While multicenter, the training set (70 AD cases) is relatively small for deep learning. Larger, more diverse training sets could further improve robustness.
  • Integration: The authors note that for AI to be truly effective, it must be seamlessly integrated into existing radiology workflows (PACS/RIS) to avoid adding "click fatigue" for doctors.

The authors intend to make the tool publicly available for further scientific evaluation.

Summary updated

Abstract

Early detection of aortic dissection (AD) is crucial for timely treatment. A modern convolutional neural network (CNN) achieved 93.5% sensitivity and 100.0% specificity for AD detection on multicenter, heterogeneous CT data. These results demonstrate the potential of streamlined, optimized CNNs for robust AD detection on CT, supporting fast clinical response.
BibTeX
@inproceedings{norajitra2025automated,
  author = {Tobias Norajitra and Michael Baumgartner and Lucas R. Cusumano and Jesus G. Ulloa and Caterina Rizzo and Florian Haag and Alexander Hertel and Nils Rathmann and Steffen J. Diehl and Stefan O. Schoenberg and Klaus H. Maier-Hein and Johann S. Rink},
  title = {{Automated AI detection of thoracic aortic dissection on CT imaging}},
  booktitle = {European},
  year = {2025},
  doi = {10.1186/s41747-025-00640-8},
}
Figure