Background/Objectives: Endoscopic airway imaging, used in endoscopy-guided airway management, enables geometry-based assessment of the airways. However, COCO-format datasets often include fragmented regions and geometrically inconsistent polygon annotations. Such inconsistencies may reduce reproducibility and spatial stability in deep learning-based image segmentation. This study proposes a systematic annotation quality-control (QC) workflow to improve dataset integrity before model training. Methods: The phantom subset of the Upper Airway Anatomical Landmark (UAAL) dataset containing 4526 polygon instances across 2746 frames (2267 training; 479 validation) was analyzed. The QC pipeline validated polygon structure, generated masks at native image resolution, and removed small noise-like instances using an area threshold. A YOLOv8-seg model was trained using (i) original annotations and (ii) QC-refined annotations. Performance was evaluated using precision, recall, mAP@0.5, mAP@0.5:0.95, and Dice similarity coefficient (DSC). Frame-level DSC values were compared using the Wilcoxon signed-rank test. Results: Annotation QC improved boundary consistency and reduced mask fragmentation. Training with QC-refined annotations increased box mAP@0.5:0.95 from 0.602 to 0.628 and mean DSC from 0.823 to 0.830 (p < 0.05). A pilot evaluation on the UAAL clinical subset also showed improved performance, with Box mAP@0.5 increasing from 0.635 to 0.706 and Mask mAP@0.5 from 0.631 to 0.704. Conclusions: Annotation-level QC enhances segmentation robustness without modifying the network architecture. The proposed workflow improves the reproducibility of results in endoscopic image segmentation and may support more stable geometry-based airway analysis in deep learning applications.
Background: Unexpected difficult laryngoscopy remains one of the leading causes of life-threatening complications in anesthetic practice. Conventional airway assessment tests (e.g., Mallampati classification, thyromental distance) are inherently subjective and demonstrate limited diagnostic accuracy. We aimed to develop a hybrid artificial intelligence model that extracts three-dimensional geometric ratios from standardized facial photographs using deep learning and graph-based modeling to predict difficult laryngoscopy. Methods: In this prospective observational study, consecutive patients scheduled for elective surgery were enrolled. Preoperatively, standardized facial photographs were obtained from eight projection angles. Instead of raw pixel-based image analysis, the images were transformed into three-dimensional anatomical graph structures following local illumination normalization using Contrast Limited Adaptive Histogram Equalization (CLAHE), combined with MediaPipe Face Mesh (468 landmark points) and Pose Estimation (33 skeletal key points). Clinically relevant anatomical indices, including Retrognathia, Overjet, Neck Bulk, and Airway Visibility, were mathematically derived from these coordinates and combined with routinely collected clinical and anthropometric variables as structured model inputs. Cormack–Lehane (CL) grading was used as the reference standard (Grades I–II: non-difficult laryngoscopy; Grades III–IV: difficult laryngoscopy). The structured dataset was used to train XGBoost, Random Forest, and Tabular Transformer models incorporating multi-head attention. To prevent data leakage, Synthetic Minority Over-sampling Technique (SMOTE) augmentation and scaling were applied exclusively to the training set. Results: Preoperative images and anatomical ratios of 198 patients were analyzed. Among the derived indices, the Mouth Opening Ratio and Oropharyngeal Visibility Index showed the highest landmark acquisition success, with extraction in 193 of 198 patients (97.5%). The proposed hybrid artificial intelligence (AI) model achieved a test-set accuracy of 80.0%, sensitivity (recall) of 85.7%, and an area under the curve (AUC) of 0.808 for difficult laryngoscopy prediction. Conclusions: Unlike conventional convolutional neural networks (CNNs) relying on raw pixel processing, the proposed model combines standardized pixel-space measurements and normalized geometric ratios. This landmark-based representation may reduce sensitivity to some acquisition-related variability under standardized imaging conditions. These findings support the feasibility of a landmark-based, geometry-driven approach for objective preoperative airway risk assessment. However, external validation and direct comparison with standardized bedside airway assessment are required before routine clinical implementation.