Non-Contrast CT-Based Deep Learning Workflow for Aortic Dissection Detection

A multicenter retrospective study from 6 hospitals developed a deep learning workflow for screening aortic dissection using non-contrast chest CT scans. The model-development dataset includes 458 patients, yielding 17,484 aortic image slices, with an external evaluation cohort of 74 patients from 2 additional hospitals. The workflow, authored by Jie Hong and last updated in April 2026, achieved areas under the ROC curve of up to 0.801 for binary detection and 0.796 for Stanford type classification.

Use Cases

• Training a binary classifier for aortic dissection detection based on non-contrast CT image slices.
• Developing a model for Stanford type A versus type B classification based on extracted aortic region-of-interest features.
• Implementing a triage support system for regional hospitals based on the described deep learning workflow.
• Analyzing model interpretability via gradient-weighted class activation mapping focusing on intimal flaps and lumen interfaces.

Strengths

• Dataset includes 458 patients from 6 hospitals, with a clear split into training (n=358), validation (n=50), and test (n=50) cohorts.
• External validation was performed on an independent cohort of 74 patients from 2 held-out hospitals.
• Model performance metrics are provided, including AUCs of 0.801 and 0.786 for detection and specific accuracy, sensitivity, and specificity for type classification.

Limitations

• Row count is unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.
• Data may reflect geographic or institutional bias inherent to the specific regional medical community studied.

Provenance

Source

figshare, authored by Jie Hong.

Collection Method

Multicenter retrospective study collecting non-contrast chest CT scans.

Time Range

null

Freshness

Last updated 2026-04-22 14:21:11; freshness should be verified.

Geography

Regional medical community (specific region not named).