[Paper Review] OralBBNet: Spatially Guided Dental Segmentation of Panoramic X-Rays with Bounding Box Priors

Devichand Budagam, Azamat Zhanatuly Imanbayev+4 more|arXiv (Cornell University)|Jun 6, 2024

Dental Radiography and Imaging4 citations

TL;DR

This paper proposes OralBBNet, a novel two-stage deep learning framework that integrates YOLOv8 for bounding box detection with a modified U-Net (BB-UNet) for instance segmentation of teeth in panoramic X-rays. By leveraging spatial priors from YOLOv8 predictions, the method achieves a 10–15% improvement in dice coefficient over standard U-Net and a 3% increase in mAP for teeth classification, demonstrating superior performance in complex dental structures like molars and premolars.

ABSTRACT

Teeth segmentation and recognition play a vital role in a variety of dental applications and diagnostic procedures. The integration of deep learning models has facilitated the development of precise and automated segmentation methods. Although prior research has explored teeth segmentation, not many methods have successfully performed tooth segmentation and detection simultaneously. This study presents UFBA-425, a dental dataset derived from the UFBA-UESC dataset, featuring bounding box and polygon annotations for 425 panoramic dental X-rays. In addition, this paper presents the OralBBNet architecture, which is based on the best segmentation and detection qualities of architectures such as U-Net and YOLOv8, respectively. OralBBNet is designed to improve the accuracy and robustness of tooth classification and segmentation on panoramic X-rays by leveraging the complementary strengths of U-Net and YOLOv8. Our approach achieved a 1-3% improvement in mean average precision (mAP) for tooth detection compared to existing techniques and a 15-20% improvement in the dice score for teeth segmentation over state-of-the-art (SOTA) solutions for various tooth categories and 2-4% improvement in the dice score compared to other SOTA segmentation architectures. The results of this study establish a foundation for the wider implementation of object detection models in dental diagnostics.

Motivation & Objective

• Address the challenge of accurate, simultaneous teeth instance segmentation and classification in panoramic X-rays, which are often low-contrast and noisy.
• Overcome limitations of prior methods that treat all teeth as a single class or fail to generalize across complex anatomical variations.
• Leverage spatial prior knowledge (bounding boxes) to enhance U-Net’s segmentation accuracy, especially in dense or degraded tooth regions.
• Develop a robust, efficient pipeline suitable for clinical deployment by combining fast inference from YOLOv8 with precise segmentation from BB-UNet.
• Create a new annotated dental dataset with polygon and bounding box annotations to support advanced instance segmentation and classification tasks.

Proposed method

• Utilize YOLOv8 as a one-stage object detector to predict bounding boxes for individual teeth in panoramic X-rays, providing spatial priors.
• Integrate YOLOv8’s predicted bounding boxes into the skip connections of a U-Net architecture to guide feature learning and improve localization accuracy.
• Train the BB-UNet model end-to-end using cross-entropy and Dice loss to optimize pixel-wise segmentation performance.
• Apply data augmentation and class balancing techniques to improve generalization, especially for rare classes like implants or severely degraded teeth.
• Use transfer learning and fine-tuning to initialize the backbone networks, enhancing convergence and performance on limited dental data.
• Evaluate the pipeline on a newly created dataset of 425 panoramic X-rays with both bounding box and polygon annotations for precise evaluation.

Figure 1: FDI Teeth Numbering Notation

Experimental results

Research questions

• RQ1 Can integrating bounding box priors from YOLOv8 significantly improve the accuracy of U-Net-based instance segmentation in panoramic dental X-rays?
• RQ2 How does the performance of BB-UNet compare to standard U-Net and other state-of-the-art models like Mask R-CNN and PANet in teeth segmentation across different tooth types?
• RQ3 To what extent does the quality of bounding box predictions from YOLOv8 affect the final segmentation performance of BB-UNet?
• RQ4 Does the proposed method maintain high performance on challenging cases such as teeth with fillings, implants, or severe degradation?
• RQ5 Can a lightweight, one-stage detection model like YOLOv8 enable faster inference while maintaining high segmentation accuracy in clinical dental imaging?

Key findings

• BB-UNet achieved a 10–15% improvement in dice coefficient over standard U-Net across all tooth categories, with the highest gains observed in molars and premolars.
• The model achieved a 3% increase in mAP for teeth classification compared to existing methods, demonstrating superior classification performance.
• YOLOv8 outperformed Mask R-CNN, HTC, and ResNeSt in mAP (72.9%) and matched or exceeded them in AP50 (94.6%), especially on complex, diverse test sets.
• BB-UNet's performance was highly dependent on accurate YOLOv8 predictions; when bounding boxes were missing, performance degraded to U-Net levels, confirming the reliance on spatial priors.
• The model achieved comparable segmentation performance to Mask R-CNN and U-Net + Mask R-CNN but with faster inference, making it more suitable for real-time clinical use.
• The newly created dataset of 425 panoramic X-rays with both bounding box and polygon annotations enables more accurate evaluation and future benchmarking in dental instance segmentation.