A nnU-Net based deep learning approach for detection and positional classification of maxillary impacted canines in cone beam computed tomography images: A pilot study

Sažetak

Background: Maxillary impacted canines represent one of the most clinically significant eruption disturbances due to their association with severe complications such as root resorption and complex orthodontic treatment requirements. Accurate three-dimensional localization is therefore essential for diagnosis and treatment planning. The aim of this study was to develop and evaluate a deep learning-based automated decision support system using a nnU-Net architecture for the detection, segmentation and positional classification of impacted maxillary canines on cone-beam computed tomography (CBCT) images. Materials and Methods: This retrospective study included 49 CBCT datasets obtained from the archive of the Department of Oral and Maxillofacial Radiology. All datasets were anonymized and converted from DICOM to NIfTI format. A total of 44 datasets were used for training and 5 for testing. Manual segmentation of impacted canines was performed using CranioCatch annotation software (CranioCatch, Eskişehir, Türkiye) to establish ground-truth annotations. The nnU-Net v2 architecture was used for model training and automatic segmentation. Model performance was evaluated using the Dice similarity coefficient, Jaccard index, Hausdorff distance, and confusion matrix-based metrics, including precision, recall, and F1 score. Results: The model demonstrated variable performance across different classes. The highest segmentation performance was observed in palatal impactions (Dice = 0.685, recall = 0.744), whereas horizontal and vestibular impactions showed poor segmentation (Dice = 0). Despite very high overall accuracy values (>0.999), high false negative rates were observed in underrepresented classes, indicating significant class imbalance. Conclusion: The nnU-Net model demonstrated promising performance in detecting and segmenting palatal impacted canines; however, its performance was limited in other classes due to dataset imbalance and anatomical variability. Future studies with larger and more balanced datasets are required to improve clinical applicability.