A Unified Deep Learning Framework for Visual Diagnosis of Palatal Radicular Grooves in CBCT Scans: A Multicenter...

Palatal radicular grooves (PRGs) posed diagnostic challenges due to their complex root anatomy and subtle manifestations in cone-beam computed tomography (CBCT). This study aimed to develop a deep learning framework for the automated three-dimensional visualization, diagnosis, and classification of PRG lesions.

A unified framework (PRG-Net) integrating tooth segmentation, PRG diagnosis, and lesion classification was developed. A retrospective multicenter diagnostic accuracy study was conducted using CBCT datasets with varying fields of view from one internal validation site and 3 external centers to evaluate generalizability and performance for segmentation, diagnosis, and classification tasks. The impact of PRG-Net on dentists' diagnostic accuracy, classification consistency, and workflow efficiency was also assessed.

PRG-Net demonstrated strong generalizability across all datasets. For tooth segmentation, it achieved a mean Dice similarity coefficient of 97.1% [95% CI: 96.4, 97.7]. Diagnostic performance yielded area under the curve of 94.4% (internal) and 85.2%-90.0% (external). Classification area under the curve were 91.4% [95% CI: 86.8, 96.1] for Type I, 88.5% [95% CI: 81.1, 95.8] for Type II, and 96.9% [95% CI: 91.6, 100] for Type III, with consistent cross-center reproducibility. In clinical validation, PRG-Net significantly improved dentists' diagnostic accuracy and inter-rater classification agreement while substantially reducing interpretation time.

PRG-Net provided a robust, automated solution for PRG assessment in CBCT. It facilitated earlier and more precise diagnosis, improved inter-rater reliability, and streamlined workflow, demonstrating strong potential as a clinically valuable decision-support tool to guide treatment planning and improve patient outcomes.