TI-YOLO: A Lightweight and Efficient Anatomical Structure Detection Model for Tracheal Intubation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate glottis visualization is crucial for safe tracheal intubation (TI), but video laryngoscopy faces limitations in field of view and computational power.
• Existing deep learning (DL) models often fail to balance high accuracy with real-time clinical deployment needs, especially in difficult airway scenarios.

Purpose of the Study:

• To develop a lightweight and efficient object detection model for real-time glottis identification during TI.
• To improve the accuracy and robustness of anatomical structure detection in challenging clinical settings.

Main Methods:

• Proposed TI-YOLO, a lightweight object detection model based on YOLOv11, incorporating Bidirectional Feature Pyramid Network (BiFPN) for multi-scale feature fusion and Deformable Attention Transformer (DAT) for enhanced perception.
• Optimized the backbone using MobileNetV4 and employed the Slide Weight Function (SWF) loss to address class imbalance.
• Validated on a custom dataset and evaluated on an embedded platform (OrangePi 5).

Main Results:

• TI-YOLO achieved a mean Average Precision at IoU 0.50 (mAP50) of 0.902, a 3.8% improvement over YOLOv11.
• Reduced computational load by 10.5% (FLOPs) and parameters by 28.9%, with a model size of 4.6 MB.
• Real-time inference speed exceeded 50 frames per second (FPS) on an embedded platform, meeting clinical requirements.

Conclusions:

• TI-YOLO offers a significant advancement in AI-powered glottis detection for tracheal intubation, balancing high accuracy with computational efficiency.
• The model's lightweight design and real-time performance make it suitable for clinical deployment, potentially enhancing patient safety and procedural success in difficult airways.