Study: YOLO11m Shows Better Generalization Than YOLOv8m on Nasal CBCT Detection
Deep-learning models can now detect three key nasal obstructions on cone-beam CT scans with high accuracy, a study reported.
Researchers trained YOLOv8m and the newer YOLO11m on 3,708 augmented CBCT slices from 120 patients scanned between 2023 and 2024 using NewTom GIANO (Verona, Italy) and Morita 3D Accuitomo 170 (Kyoto, Japan) devices.
The study evaluated automatic detection of concha bullosa, nasal septal deviation, and the maxillary sinus ostium — structures associated with osteomeatal complex obstruction. The dataset comprised 618 CBCT slices per structure (1,854 total), labeled for concha bullosa and nasal septal deviation presence and maxillary sinus ostium status (open or closed). Following horizontal flipping for data augmentation, the images were divided into training (80%), validation (10%), and testing (10%) sets.
Both models achieved approximately 99% detection performance for concha bullosa, with test mean average precision at the 0.5 intersection-over-union threshold ([email protected]) of roughly 0.99. For maxillary sinus ostium detection, comparable training performance was observed ([email protected] ≈ 0.98), while YOLO11m achieved a higher test [email protected] of 0.943 compared with YOLOv8m’s 0.929.
In nasal septal deviation detection and localization, YOLO11m demonstrated favorable performance, with test [email protected] reaching 0.995 against YOLOv8m’s 0.987. Comparable findings were observed across both CBCT devices. Overall, YOLO11m demonstrated better generalization performance, whereas YOLOv8m was more time-efficient.
Feasibility study tests contrast enhancement on 203 scans
Standardized coronal images were obtained from 83 scans with concha bullosa and 119 without, then preprocessed through a hybridized contrast-enhancement method using discrete wavelet transform. The images were split into 162 training and 41 testing cases.
A separate feasibility study published earlier this year assessed deep-learning detection of concha bullosa alone, using 203 CBCT scans from a dental teaching hospital and a hybridized contrast-enhancement method with discrete wavelet transform. That study reported the CE-DWT-ResNet101-RF model achieved an accuracy of 91.7% and an area under the curve of 98%, with precision, recall, and F1 score all at 92%.