基于双域分析与自适应注意力的牙龈炎症分级智能模型
本研究提出HFSA-DETR模型,融合空间-频率双域分析与自适应稀疏注意力,用于牙龈炎症分级。内部数据集显示精度、召回率和mAP50分别较基线提升3.9%、3.8%和3.3%;外部验证集精度92.8%、召回率91.5%、mAP50 93.2%。模型参数与计算量显著降低,有助于临床早期检测牙龈炎症并提供量化依据。
要点
- HFSA-DETR在外部验证集实现92.8%精度、91.5%召回率和93.2% mAP50。
- 内部数据集上精度、召回率、mAP50较基线分别提升3.9%、3.8%、3.3%。
- 模型参数与计算需求显著降低,适合资源有限的口腔诊所。
核心结论
- HFSA-DETR通过融合空间-频率双域分析与自适应稀疏注意力机制,实现牙龈炎症的准确分级。
- 该模型在保持高分级精度的同时显著降低计算成本,为牙龈炎症的客观定量评估提供有效技术方案。
研究设计与样本
- 研究纳入2300张牙龈炎症图像用于模型训练,并使用包含1096张图像的开源数据集评估泛化能力。
- 在RT-DETR骨干网络中引入空间-频率混合变换模块,在颈部网络引入自适应稀疏自注意力机制。
主要发现
- 在自建内部牙龈炎分级数据集上,HFSA-DETR的精度、召回率和mAP50分别较基线模型提升3.9%、3.8%和3.3%。
- 模型参数数量和计算需求显著降低。
- 在外部验证数据集上,模型达到92.8%精度、91.5%召回率和93.2% mAP50,在不同临床成像条件下表现出优异泛化能力。
临床怎么用
- 该模型可帮助牙医早期检测和干预牙龈炎症,为临床决策提供定量依据。
- 尤其适用于资源有限的口腔诊所。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
OBJECTIVE: This study proposes the gingival inflammation grading model Hybrid-Frequency Self-Attention Detection Transformer (HFSA-DETR), which is based on a dual-domain spatial-frequency analysis and adaptive sparse attention mechanism. The model aims to improve the grading accuracy of subtle inflammatory features and the sensitivity of identifying early gingival inflammation under non-uniform lighting conditions. METHODS: This study included 2300 gingival inflammation images for model training and evaluated the model's generalisation using an open-source dataset containing 1096 images of gingival inflammation. A spatial-frequency hybrid transformation module was introduced into the Real-Time Detection Transformer (RT-DETR) backbone network. Through dual-domain collaborative analysis, it simultaneously captured the colour, morphological, and texture periodic characteristics of gingival inflammation, addressed non-uniform lighting, and enhanced the model's ability to perceive subtle changes in gingival inflammation. Introducing an adaptive sparse self-attention mechanism in the neck network enabled the model to adaptively focus on key areas of gingival inflammation, reducing computational complexity while maintaining grading accuracy. RESULTS: On the self-built internal gingivitis grading dataset, the HFSA-DETR model demonstrated improvements of 3.9%, 3.8%, and 3.3% in precision, recall, and mean average precision (mAP)50, respectively, compared to the baseline model; it exhibited significantly reduced parameter counts and computational requirements. On external validation datasets, the model achieved 92.8% precision, 91.5% recall, and 93.2% mAP50, demonstrating exceptional generalisation across diverse clinical imaging conditions. CONCLUSION: The HFSA-DETR model achieved accurate grading of gingival inflammation by fusing spatial-frequency dual-domain analysis with an adaptive sparse attention mechanism. This model significantly reduces computational costs while maintaining high grading accuracy, providing an effective technical solution for objective quantitative assessment of gingival inflammation. CLINICAL SIGNIFICANCE: The HFSA-DETR model proposed in this study can help dentists detect and intervene early in gingival inflammation, providing a quantitative basis for clinical decision-making, especially for dental clinics with limited resources.
引用
- Chen M, Dai Y, Hu P, et al. A novel intelligent gingival inflammation grading model based on dual-domain analysis and adaptive attention. J Dent. 2026. PMID: 41687952. doi:10.1016/j.jdent.2026.106570
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。