Transitioning from a Single-Model Baseline to a Multi-Stage Deep Learning Pipeline: A Head-to-Head Evaluation for Tooth Numbering in Intraoral Photographs

Despite the growing use of deep learning (DL) in dental imaging, no study has compared the relative effectiveness of single-stage and multi-stage DL architectures for World Dental Federation (FDI)-based tooth numbering. Therefore, in this paper, we present a head-to-head benchmarking study comparing a single-stage baseline against a multi-stage pipeline for automated tooth detection and the FDI numbering system. The study was conducted on a dataset of 3,600 intraoral occlusal photographs. The baseline method uses a 28-class You Only Look Once version 8 medium (YOLOv8m) object detector to directly predict per tooth labels from annotated images. On the other hand, the pipeline method leverages a multi-stage process in which each image is first classified by a Residual Neural Network classifier (ResNet-18) to identify the arch as maxillary or mandibular. As the second a YOLOv8m is applied to localize the teeth, where each resulting bounding box is forwarded to the appropriate ResNet-101 classifiers to assign an anatomical class. Finally, a deterministic mapping function is used for integrating the anatomical label, tooth position, and the arch type to generate the corresponding FDI number. Both approaches are trained on identical data splits and evaluated under the same matching protocol and decision thresholds finetuned on a held-out validation set. Evaluation at Intersection over Union (IoU) ≥ 0.5 and confidence ≥ 0.7 resulted in the following: accuracy, precision, recall, and F1-score values of 0.871, 0.879, 0.873, and 0.875 for the baseline, and 0.948, 0.950, 0.948, and 0.949 for the pipeline, respectively. McNemar’s test yielded p=0.0002. Beyond empirical benchmarking, our study addresses a fundamental question in the literature, can a streamlined baseline achieve higher accuracy than a staged pipeline, or whether data simplification inherently enhance or compromise performance? The results provide an evidence-based and reproducible template for future dental artificial intelligence (AI) research.