CHSJ

Curr Health Sci J, vol. 52, no. 3, 2026

AI-Assisted Tooth Detection and Segmentation Workflow on Orthopantomograms Using the FDI Notation

[Original Paper]

D.A. LUTESCU(1), C.C. MANOLE(2), R. ILINCA(1), R.C.W. CHAU(3), A.C. DIDILESCU(4)


(1)Department of Medical Informatics and Biostatistics, Faculty of Dentistry,“Carol Davila” University of Medicine and Pharmacy, Bucharest, Romania;
(2)Department of Biophysics, Faculty of Dentistry, ”Carol Davila” University of Medicine and Pharmacy, Bucharest, Romania;
(3)Faculty of Dentistry, The University of Hong Kong, Hong Kong;
(4)Department of Embryology and Microbiology, Faculty of Dentistry, ”Carol Davila” University of Medicine and Pharmacy, Bucharest


Abstract:

Background: Automated tooth detection and segmentation on orthopantomograms (OPGs) is a foundational task for AI-based dental diagnostics. Most published studies have focused on subsets of the dentition or employed rectangular bounding boxes that inadequately capture overlapping dental structures. This study developed and evaluated a YOLO26s-seg instance segmentation model for polygon-based detection and FDI-based identification of all 32 permanent tooth classes on a publicly available panoramic dataset. Methodology: A convenience sample of 282 OPGs from the Tufts Dental Database was annotated with polygon-based FDI labels using an in-house browser-based platform. The dataset was partitioned at the patient level into training (n=197), validation (n=42), and test (n=43) subsets. A YOLO26s-seg model was trained using transfer learning at 1280×1280 resolution with laterality-preserving augmentation. Results: On the held-out test set, the model achieved box-level precision of 0.883 and recall of 0.892, mAP50 of 0.896 (95% CI: 0.865-0.923) for detection and 0.890 (95% CI: 0.858-0.919) for segmentation, and mAP50-95 of 0.620 and 0.556, respectively. Twenty-eight of 32 classes exceeded a box-level mAP50 of 0.80, with three of the four lowest-performing classes located in the mandibular left quadrant. Conclusions: YOLO26s-seg achieved strong mAP50 performance for instance segmentation and FDI identification of all 32 permanent tooth classes. The workflow provides a transparently reported research benchmark; clinical effectiveness was not evaluated.


Keywords: Orthopantomography, instance segmentation, deep learning, YOLO26, tooth detection, FDI numbering



Corresponding: Dan-Adrian Lutescu, Department of Medical Informatics and Biostatistics, Faculty of Dentistry, "Carol Davila" University of Medicine and Pharmacy, Bucharest, Romania, e-mail: dan-adrian.lutescu@drd.umfcd.ro


DOI 10.12865/CHSJ.52.03.09 - Download PDF
current issue cover Download cover
Download contents

Journal archive