Scholarlink

<p>Dental caries is a common oral disease that needs early and proper diagnosis to avoid severe complications. Older methods of diagnosis are compromised by subjective interpretation and limited automation, resulting in inconsistent results. To overcome these issues, this paper introduces an improved VGG 16 (ImVGG-16) deep learning-based approach to the automatic classification of dental caries from intraoral images. The system proposed herein combines several innovations in four key phases. During the preprocessing stage, an Improved Wiener Filtering technique is applied to restore image quality and eliminate noise, using a dynamically tuned Laplacian of Gaussian (LoG) filter that sharpens prominent edges and adjusts to diverse image properties. In the segmentation stage, an Improved TransU-Net model is utilized to strictly separate carious areas. The proposed model integrates Improved Residual Blocks to address vanishing gradient issues and a Multi-scale Attention Gate Mechanism to raise global context perception, allowing for accurate boundary location. In feature extraction, the system deploys an Improved Local Gabor Directional Pattern (ImLGDP) to extract dense directional texture information even in low-contrast or planar areas. This is further complemented with Median Ternary Pattern (MTP) for increased local contrast and HOG-based shape features that give a comprehensive description of the lesion. Lastly, in the classification stage, an Improved VGG-16 model is proposed. The model incorporates a custom Weighted Average Top-k Response with Batch Normalization (W-Avg-TopK-BN) mechanism that improves feature selection, stabilizes training, and reduces overfitting. By integrating these improved modules, the proposed framework generates highly accurate, robust, and generalizable outcomes, proving itself to be an effective and powerful clinical tool for automatic dental caries diagnosis. The Improved VGG-16 consistently surpasses the conventional methods with greater accuracy at 0.958, recall at 0.935 and specificity at 0.968, respectively.</p> Graphical Abstract <p></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

Improved VGG 16 with TransU-Net Segmentation for Dental Caries Classification

  • 期刊论文

  • 发表日期：

    2026年7月25日

查看全文
  • Prajakta Shinde,
  • Priyanka Paygude

摘要

Dental caries is a common oral disease that needs early and proper diagnosis to avoid severe complications. Older methods of diagnosis are compromised by subjective interpretation and limited automation, resulting in inconsistent results. To overcome these issues, this paper introduces an improved VGG 16 (ImVGG-16) deep learning-based approach to the automatic classification of dental caries from intraoral images. The system proposed herein combines several innovations in four key phases. During the preprocessing stage, an Improved Wiener Filtering technique is applied to restore image quality and eliminate noise, using a dynamically tuned Laplacian of Gaussian (LoG) filter that sharpens prominent edges and adjusts to diverse image properties. In the segmentation stage, an Improved TransU-Net model is utilized to strictly separate carious areas. The proposed model integrates Improved Residual Blocks to address vanishing gradient issues and a Multi-scale Attention Gate Mechanism to raise global context perception, allowing for accurate boundary location. In feature extraction, the system deploys an Improved Local Gabor Directional Pattern (ImLGDP) to extract dense directional texture information even in low-contrast or planar areas. This is further complemented with Median Ternary Pattern (MTP) for increased local contrast and HOG-based shape features that give a comprehensive description of the lesion. Lastly, in the classification stage, an Improved VGG-16 model is proposed. The model incorporates a custom Weighted Average Top-k Response with Batch Normalization (W-Avg-TopK-BN) mechanism that improves feature selection, stabilizes training, and reduces overfitting. By integrating these improved modules, the proposed framework generates highly accurate, robust, and generalizable outcomes, proving itself to be an effective and powerful clinical tool for automatic dental caries diagnosis. The Improved VGG-16 consistently surpasses the conventional methods with greater accuracy at 0.958, recall at 0.935 and specificity at 0.968, respectively.

Graphical Abstract