AKUDENTAL teeth instance segmentation dataset: a cross-dataset analysis

TL;DR

The paper introduces the AKUDENTAL dataset for dental radiograph segmentation and shows that annotation differences across datasets significantly affect AI model performance.

Contribution

The novel contribution is the AKUDENTAL dataset with annotated dental structures and a cross-dataset analysis highlighting annotation inconsistencies as a key challenge.

Findings

Performance differences in AI models are largely due to variations in annotation protocols across datasets.

Mean Average Precision scores varied widely, from 0.34 on DENTEX to 0.71 on Dual-labeled datasets.

Annotation inconsistencies are a critical barrier to developing universally applicable dental AI models.

Abstract

Artificial Intelligence (AI) is reshaping diagnostics and disease prevention in the dental domain. Panoramic X-ray imaging is central to this progress but demands large, high-quality annotated datasets. We therefore present AKUDENTAL, a new dataset for instance segmentation of dental radiographs, to serve as a resource for model development and to assess the challenges of generalizability.

We annotated 333 panoramic images, labeling 9,956 structures across 32 individual teeth and three restorative categories: implants, bridges, and crown–filling. We established semantic segmentation, object detection, and instance-segmentation baselines using UNet, DeepLabV3 + , YOLOv11, and Mask R-CNN models. Generalizability was assessed via 5-fold cross-validation and a cross-dataset evaluation on the Tufts, DENTEX, and Dual-labeled datasets.

A cross-dataset evaluation on the Tufts, DENTEX, and Dual-labeled datasets revealed that variations in annotation protocols are a significant factor contributing to performance differences. The cross-dataset evaluation demonstrated widely varying performance, with mean Average Precision (mAP) scores for multiclass detection ranging from a low of 0.34 on the DENTEX dataset to 0.71 on the Dual-labeled dataset Our analysis illustrates how such discrepancies can impact the interpretation of model performance.

The AKUDENTAL dataset provides a robust new resource for the field. The performance disparities revealed in our cross-dataset analysis are not model limitations but instead strengthen the argument that annotation inconsistencies are a critical barrier to developing universally applicable AI. This highlights the imperative for broader standardization in data annotation, extending beyond tooth identification to encompass common dental procedures and restorations.

Figures