Deep Learning Application of YOLOv8 for Aortic Dissection Screening Using Non-contrast Computed Tomography

Area of Science:

• Artificial Intelligence in Medical Imaging
• Cardiovascular Radiology
• Deep Learning for Diagnostic Support

Background:

• Acute aortic dissection (AD) is a critical condition requiring rapid diagnosis.
• Non-contrast computed tomography (CT) is utilized for AD detection, but radiologist interpretation can be variable and time-consuming.
• There is a need for improved, efficient, and accurate diagnostic tools for AD.

Purpose of the Study:

• To develop and validate an interpretable YOLOv8 deep learning model for detecting AD on non-contrast CT scans.
• To compare the diagnostic performance and efficiency of the AI model against radiologists.

Main Methods:

• A retrospective study involving 1,138 non-contrast CT scans from five institutions, split into training, internal, and external validation cohorts.
• An interpretable YOLOv8 deep learning model was trained on annotated CT images.
• Performance was assessed using AUC, sensitivity, specificity, and inference time, with comparisons to vascular interventional radiologists, general radiologists, and residents. Grad-CAM was used for interpretability.

Main Results:

• The YOLOv8s model demonstrated high diagnostic performance with an AUC of 0.970 in the external validation cohort.
• The model's sensitivity (0.976) and specificity (0.935) were comparable or superior to radiologists.
• The model's inference time (3.47 seconds) was significantly faster than the mean radiologist interpretation time (25.32 seconds).

Conclusions:

• The YOLOv8s deep learning model reliably detects AD on non-contrast CT, surpassing radiologists in speed and accuracy.
• This AI tool offers potential for enhancing AD screening, aiding clinical decision-making, and improving diagnostic quality.
• The model's interpretability, confirmed by Grad-CAM, supports its clinical utility.