Scholarlink

<p>Dental caries is one of the most prevalent oral diseases worldwide and remains a major cause of tooth deterioration and loss when not detected at an early stage. Accurate segmentation of carious lesions in panoramic dental radiographs is challenging due to low lesion contrast, image noise, anatomical complexity, and the small size of affected regions. To address these challenges, this study proposes <b>AST_UNet</b>, an <b>Attention Swin Transformer U-Net framework for automated dental caries segmentation.</b> The proposed framework integrates Wiener Filter-based denoising and Dynamic Histogram Equalization (DHE)-based contrast enhancement to improve image quality before segmentation. Subsequently, a Swin Transformer encoder is employed to capture hierarchical local and global contextual features, while an attention-guided U-Net decoder facilitates accurate lesion localization and boundary delineation. Experiments were conducted on a publicly available panoramic dental radiograph dataset using standardized training, validation, and testing protocols. The proposed AST_UNet achieved a Dice Similarity Coefficient (DSC) of 97.86%, Intersection over Union (IoU) of 95.81%, Pixel Accuracy of 99.98%, Precision of 99.43%, and recall of 96.34%. Comparative evaluation against SegNet, U-Net, Attention U-Net, and SwinUNet demonstrated consistent performance improvements. Ablation studies further confirmed the contributions of the Swin Transformer encoder, Wiener Filter, and Dynamic Histogram Equalization to the overall segmentation performance. Although additional validation on larger multi-center datasets is required, the results indicate that AST_UNet provides an effective and reliable framework for automated dental caries segmentation in panoramic radiographs and may support future computer-aided dental diagnostic systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

AST_UNet: attention swin transformer U-net-based dental panoramic caries segmentation for early diagnosis

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月5日

查看全文
  • Sriramoju Archana,
  • C. Madan Kumar,
  • M. Brindha

摘要

Dental caries is one of the most prevalent oral diseases worldwide and remains a major cause of tooth deterioration and loss when not detected at an early stage. Accurate segmentation of carious lesions in panoramic dental radiographs is challenging due to low lesion contrast, image noise, anatomical complexity, and the small size of affected regions. To address these challenges, this study proposes AST_UNet, an Attention Swin Transformer U-Net framework for automated dental caries segmentation. The proposed framework integrates Wiener Filter-based denoising and Dynamic Histogram Equalization (DHE)-based contrast enhancement to improve image quality before segmentation. Subsequently, a Swin Transformer encoder is employed to capture hierarchical local and global contextual features, while an attention-guided U-Net decoder facilitates accurate lesion localization and boundary delineation. Experiments were conducted on a publicly available panoramic dental radiograph dataset using standardized training, validation, and testing protocols. The proposed AST_UNet achieved a Dice Similarity Coefficient (DSC) of 97.86%, Intersection over Union (IoU) of 95.81%, Pixel Accuracy of 99.98%, Precision of 99.43%, and recall of 96.34%. Comparative evaluation against SegNet, U-Net, Attention U-Net, and SwinUNet demonstrated consistent performance improvements. Ablation studies further confirmed the contributions of the Swin Transformer encoder, Wiener Filter, and Dynamic Histogram Equalization to the overall segmentation performance. Although additional validation on larger multi-center datasets is required, the results indicate that AST_UNet provides an effective and reliable framework for automated dental caries segmentation in panoramic radiographs and may support future computer-aided dental diagnostic systems.