Intelligent Recognition Network for Difficult Airway Based on Multi-View Input and Attention Aggregation.

In anesthesiology, difficult airway (DA) poses significant safety risks. Current deep learning-based DA recognition methods exhibit limited performance due to single-view reliance or simplistic multi-view fusion. To address these constraints, we propose an intelligent DA recognition network using multi-view input with attention aggregation, aimed at achieving automatic DA assessment through multi-view patient image integration. Compared with existing DA recognition techniques, our method offers several advantages. First, our network utilizes five patient images captured from different perspectives, systematically capturing essential assessment factors: maximum mouth opening, Mallampati score, neck circumference, neck length, and neck movement, thereby providing more comprehensive information. Second, we design an attention-based feature aggregation operator to fully leverage inter-view relationships, enhancing feature extraction capabilities and improving recognition performance. Additionally, we construct a multi-view patient image dataset comprising images from 481 patients, with five images per patient from distinct perspectives. Through five-fold cross-validation experiments, our method achieves accuracy of 93.54%, sensitivity of 83.33%, specificity of 96.94%, and F1-score of 86.53%, significantly outperforming other compared methods. This study was prospectively registered at the Chinese Clinical Trial Registry on August 10, 2021, with registration number ChiCTR2100049879.