《Scientific Reports》11月16日论文:DentAll 框架在1,325张全景牙片上分割 DSC 达88.21%,缺牙定位达标率由25.20%升至72.58%

Shaily Bajpai、Hamza Shafiq 与 Bumshik Lee 11月16日在《Scientific Reports》发表论文,提出名为 DentAll 的分析框架,把全景牙片中的单颗牙齿分割与缺牙位置估计整合在同一流程中。
全景牙片存在解剖变异、结构重叠与图像质量不一致等问题,临床判读长期高度依赖人工。该框架的分割模块 S-CSAB 结合扩张卷积与基于感兴趣区域的位置编码,以加强细粒度边界刻画并保持相邻牙齿间的解剖连贯性。
在此基础上的缺牙定位预测器 MTLP,不需要显式的缺牙标注,而是利用目标牙齿的身份与其同颌最近邻牙齿的几何关系,通过以患者为条件的坐标回归,估计解剖上应有的缺牙位置。
框架在1,325张全景牙片上以实例级分割指标与 COCO 风格掩膜平均精度评估。在全真值评估协议下,DSC 为88.21%,成功匹配实例中为92.74%,AP50 为97.34%,AP75 为87.20%,mAP50 至95 为68.58%。
在受控的伪缺牙定位实验中,加入上下文的 MTLP 把交并比不低于0.5的达标率从仅用索引表述时的25.20%提高到72.58%。作者表示,把实例分割与患者特异的解剖上下文结合,可为全景牙片的结构化分析提供一个计算上可行的框架,该研究获韩国国家研究基金会资助。
图中是一张牙齿全景 X 光片,DentAll 框架分析的对象正是这类全景牙片。
Scientific Reports Paper on November 16: DentAll Framework Reaches a DSC of 88.21 Percent on 1,325 Panoramic Radiographs, and Missing-Tooth Localization Success Rises From 25.20 Percent to 72.58 Percent
Shaily Bajpai, Hamza Shafiq and Bumshik Lee published a paper on November 16 in Scientific Reports introducing DentAll, a framework that integrates tooth instance segmentation in panoramic radiographs with contextual estimation of anatomically expected missing-tooth locations.
Panoramic radiographs pose persistent challenges due to anatomical variability, overlapping structures and inconsistent image quality, keeping clinical workflows heavily dependent on manual interpretation. The framework's segmentation module, S-CSAB, integrates dilated convolutions and region-of-interest-based positional encoding to enhance fine-grained boundary delineation while maintaining anatomical coherence across adjacent teeth.
Building on those features, the Missing-Tooth Localization Predictor, MTLP, estimates missing-tooth locations via patient-conditioned coordinate regression, using the target tooth's identity together with the geometry of its nearest same-arch neighboring teeth, without requiring explicit missing-tooth annotations.
DentAll was evaluated on 1,325 panoramic radiographs using instance-level segmentation metrics and COCO-style mask average precision. Under an all-ground-truth evaluation protocol, it achieved a DSC of 88.21 percent, 92.74 percent among successfully matched instances, an AP50 of 97.34 percent, an AP75 of 87.20 percent, and an mAP50 to 95 of 68.58 percent.
In controlled pseudo-missing localization experiments, the contextual MTLP improved the rate of localizations reaching an intersection-over-union of at least 0.5 from 25.20 percent with an index-only formulation to 72.58 percent. The authors say integrating instance segmentation with patient-specific anatomical context provides an effective and computationally practical framework for structured analysis of panoramic dental radiographs, and the work was supported by the National Research Foundation of Korea.
发布:GSM·美国(发布链接)
图片来源:Wikimedia Commons, File:Dental Panorama X-ray.jpg — Author: Farhang Amini. Licensing: CC BY 4.0.
GSM Global · 本篇英文版 English edition of this story

