Active learning-assisted deep learning for automated maxillary sinus segmentation in dental cone-beam computed...

Accurate automated segmentation of the maxillary sinus in cone-beam computed tomography (CBCT) scans is crucial for dental implant planning, sinus lift procedures, and pathological assessment, yet there remains a notable lack of publicly available models and datasets for this task. This study developed and evaluated a deep learning model for automated maxillary sinus segmentation using an active learning approach to minimize annotation effort while achieving high accuracy.

We utilized 249 CBCT scans from the publicly available CTooth+ and STS-Tooth datasets. Ground truth annotations were generated through semi-automated segmentation in 3D Slicer with validation by five expert dentists. An active learning strategy was employed using the nnU-Net 3D full-resolution architecture trained with five-fold cross-validation. Performance was evaluated using Dice similarity coefficient (Dice).

The model achieved a mean Dice of 0.985 (standard deviation (SD): 0.026, median: 0.992) and mean intersection over union (IoU) of 0.972 (SD: 0.044) across 249 cases. Performance remained stable across varying sinus volumes, with only rare outliers (minimum Dice: 0.712) corresponding to unusual anatomical presentations or image artifacts.

This study demonstrates that automated maxillary sinus segmentation using nnU-Net with active learning can achieve highly accurate and consistent results on diverse CBCT scans. The approach shows strong potential for clinical integration to enhance efficiency in dental treatment planning and supports the advancement of dental radiology driven by artificial intelligence (AI) through utilization of publicly available datasets.