基于深度学习的多模态气管插管智能目标检测

针对困难气道气管插管过程中内窥镜图像视角较小、目标尺度变化大、相互遮挡等问题，融合内窥镜图像和CO₂浓度信息，提出基于深度学习的多模态气管插管智能目标检测算法。首先，对传统的YOLOv3网络进行改进，利用不同扩张率的空洞卷积构建并行多分支空洞卷积模块，并对输出特征进行上采样和张量拼接；其次，根据多路CO₂浓度差异，利用矢量化定位算法定位目标中心位置，校正YOLOv3得到的边界框的中心坐标，提升小目标检测的精度，辅助气道位置的定位；最后，基于该算法，研发了新型多模态气管插管辅助装置初代样机，并在模拟气道中进行实验，验证其可行性。在模拟气道中，该新型辅助装置的操作时间中位数为15.5 s，操作成功率可达97.3%。研究结果表明，基于深度学习的多模态气管插管智能目标检测算法能够有效地辅助气管插管操作。

Multimodal intelligent target detection for endotracheal intubation based on deep learning

In order to solve the small angle of the endoscopic view, large changes of target scale and mutual occlusion during endotracheal intubation in difficult airways, a multi-modal intelligent target detection algorithm for endotracheal intubation based on deep learning was proposed with a combination of endoscopic images and carbon dioxide concentration information. First, the traditional YOLOv3 network was improved. It adopted parallel multi-branch dilated convolution block with different dilated rates to extract more information, and up-sampled the output features and concatenated these tensors. Second, the vector localization algorithm was applied to locate the center of the target according to the difference of multi-channel carbon dioxide concentration, so as to further correct the center position of the boundary box predicted by YOLOv3 network. It helped improve the accuracy of small target detection and assisted to locate airway. Finally, a prototype of the new multimodal endotracheal intubation assistant device was developed based on the proposed algorithm, and tested in a simulated airway to verify its feasibility. The effect of the prototype was satisfactory with a median operation time of 15.5 s and a success rate of 97.3%. The study shows that the multi-modal intelligent target detection algorithm for endotracheal intubation based on deep learning has good operation results and can effectively assist endotracheal intubation.