This research project aims to explore and develop fast, reliable methods for detecting aortic dissections (AD) in CT scans, tailored to emergency scenarios. To achieve this, we trained and optimized deep segmentation networks using de-identified data collected at an academic-level hospital in Germany, over a span of more than 10 years. Data preparation and annotation processes were designed to support the development and validation of the presented method. The self-configuring nnU-Net framework was utilized to optimize segmentation performance. A robust detection of AD is achieved by applying a refined thresholding rule to the segmentation output of the trained networks, effectively integrating over the segmented AD-specific structures.
- Robust and accurate detection, with performance metrics reflecting AUC greater than 0.97, sensitivity above 92%, and specificity close to 100%, validated across multiple datasets from diverse sites.
- Fully-automated and streamlined, with processing time of less than 7 minutes per case.
- Proven effectiveness in identifying AD cases that do not exhibit typical symptoms and may clinically be at risk of under-prioritization.
Our evaluation results and key findings are undergoing peer review and will be published soon. Once finalized, the trained models will be made available here for further scientific validation and collaborative research.
Stay tuned.
- The proposed tool has been trained and validated exclusively on de-identified data in a research setting and has not undergone clinical testing.
- It is not a medical product and is intended for research purposes only.
- No patient data or personal information have been published in this project.
- The automatic segmentations included in this repository and used for integration testing are based exclusively on a publicly available dataset.