A Pilot Study on YOLO26-Based Detection and Segmentation of Nasal Septum Deviation and Related Findings on CBCT

Objectives: The aim of this study was to develop and preliminarily evaluate a deep learning–based model for the automatic detection and segmentation of selected nasal structures and maxillary sinus findings on cone-beam computed tomography (CBCT) images. Material and Methods: A total of 110 expert-selected coronal CBCT slices with a wide field of view were retrospectively obtained from a larger dataset of 1000 scans. Images were preprocessed and randomly divided into training (70%), validation (15%), and test (15%) sets. A YOLOv12n-seg architecture was implemented for detection and segmentation tasks. Model performance was assessed using accuracy and area under the receiver operating characteristic curve (AUC).Results: The model achieved an accuracy of 87.5% for nasal septum deviation (NSD) and inferior concha hypertrophy (ICH). For maxillary sinus mucosal thickening (MSMT), the accuracy was 81.3%. The AUC values were 0.84 for NSD, 0.85 for ICH, and 0.68 for MSMT, indicating relatively lower performance for soft-tissue-related findings. Conclusion: The proposed deep learning model demonstrated promising performance for the detection of anatomically distinct nasal structures on selected CBCT slices. However, the limited sample size, slice-based design, and lack of external validation restrict the generalizability of the findings. This study should be considered a preliminary pilot investigation, and further research using larger, patient-based datasets and clinical comparisons is required before routine clinical application.