Real-time, high-precision detection of aortic dissection key components for emergency triage using an enhanced deep learning model.

BACKGROUND

Aortic dissection (AD) is a life-threatening cardiovascular emergency. For Type B AD (TBAD), rapid and accurate identification of the true lumen (TL), false lumen (FL), and false lumen thrombus (FLT) from CTA is critical for risk stratification and treatment planning. However, existing deep learning solutions often lack real-time capability and fail to address the detection of FLT.

PURPOSE: To develop a real-time, high-precision deep learning framework for the simultaneous detection of all three key AD components to support emergency triage.

METHODS

We propose AD-YOLO11, an enhanced YOLOv11 model integrating three key innovations: (1) a Recursive Information Distillation Network (RIDNet) for CTA noise suppression, (2) a Triplet Attention Mechanism for spatial and channel feature enhancement, and (3) the MPDIoU loss function for optimized bounding box regression. The model was trained and internally validated on a dataset of 25 176 slices from 106 TBAD patients and externally validated on 18 238 slices from 71 independent patients.

RESULTS

On internal validation, AD-YOLO11 achieved a precision of 0.991 ± 0.004, recall of 0.936 ± 0.006, mAP@0.5 of 0.951 ± 0.007, and mAP@0.5:0.95 of 0.883 ± 0.008. It maintained high performance on the external test set, demonstrating strong generalizability. The inference speed was 3.18 ± 0.23 ms per slice on GPU, and it remained clinically feasible on CPU (53.15 ± 2.76 ms per slice).

CONCLUSIONS

AD-YOLO11 achieves millisecond-level, high-accuracy detection of all three critical Type B aortic dissection components from CTA images. Its efficient inference on both GPU and CPU makes it a promising frontline tool for rapid triage in emergency and resource-limited settings, effectively complementing time-consuming 3D segmentation for aortic dissection assessment.