FD-SOS: Vision-Language Open-Set Detectors for Bone Fenestration and Dehiscence Detection from Intraoral Images

TL;DR

FD-SOS tackles detecting bone fenestration and dehiscence from intraoral images, addressing the radiation and accessibility limitations of CBCT. It introduces a multi-task vision-language open-set detector that leverages external dental semantics through conditional contrastive denoising (CCDN) and teeth-specific matching assignment (TMA). The approach achieves state-of-the-art FD detection on FDTooth/FDTOOTH datasets, surpassing dental professionals in recall at matched precision and demonstrating robustness to missing posterior annotations. This framework has practical clinical impact by enabling accurate, non-radiative screening and holds potential for extension to other dental diseases as multimodal dental data resources grow.

Abstract

Accurate detection of bone fenestration and dehiscence (FD) is crucial for effective treatment planning in dentistry. While cone-beam computed tomography (CBCT) is the gold standard for evaluating FD, it comes with limitations such as radiation exposure, limited accessibility, and higher cost compared to intraoral images. In intraoral images, dentists face challenges in the differential diagnosis of FD. This paper presents a novel and clinically significant application of FD detection solely from intraoral images. To achieve this, we propose FD-SOS, a novel open-set object detector for FD detection from intraoral images. FD-SOS has two novel components: conditional contrastive denoising (CCDN) and teeth-specific matching assignment (TMA). These modules enable FD-SOS to effectively leverage external dental semantics. Experimental results showed that our method outperformed existing detection methods and surpassed dental professionals by 35% recall under the same level of precision. Code is available at: https://github.com/xmed-lab/FD-SOS.

Figures (5)

• Figure 1: (a) Our collected dataset for FD detection focuses on anterior teeth without posterior teeth annotations, as indicated by dashed bounding boxes. (b) Publicly available datasets for teeth detection, incorporating two classes. (c) FD detection Precision-Recall Curve shows that our method outperforms other existing detection methods and even professional dentists. RandomCLS is an always-positive predictor on the true bounding boxes to establish the random average precision (AP).
• Figure 2: Framework for FD-SOS.
• Figure 3: (a) Conditional Contrastive Denoising (CCDN) improves the detection decoder by utilizing attribute-based contrastive denoising. (b) Teeth-Specific Matching Assignment (TMA) leverages positional priors to mask posterior teeth and focus on the differential diagnosis of FD in frontal intra-oral images.
• Figure 4: Qualitative results for the best-performing methods on FD detection.
• Figure 5: Teeth detection is relatively an easier task. Results on the validation set for the cross-task teeth detection, encompassing both "Anterior" and "Posterior" categories, demonstrate that open-set detectors do not require a warmup phase, as they are already familiar with the task of teeth detection from the outset, unlike other detection models such as DeformableDETR, DINO, or DiffusionDETR models. The achieved performance for teeth detection, with an $AP_{50} > 90\%$ indicates that it is effectively a solved problem in the field of computer vision.