Scholarlink

In orthodontic diagnosis and treatment planning, accurate localization of landmarks in X-ray cephalograms is essential. While deep learning approaches have shown promise in automating landmark identification, ensuring consistent performance and strong generalization remains challenging due to the lack of diverse datasets. This study presents a novel architecture based on Multi-Scale Convolutional Neural Network (CNN) and Heatmap Regression for automatic cephalometric landmark detection. This method captures global relationships among landmarks as well as the distinct characteristics of individual landmarks, allowing for discriminative analysis across multiple scales. We utilize a newly developed benchmark dataset named “Aariz”. The proposed framework attained an overall mean radial error (MRE) of 1.942 ± 3.782 on the test dataset, with the success detection rates 69.01% and 91.33% within the 2 mm and 4 mm ranges, respectively. Additionally, we compared the performance of coordinate regression and heat map regression. These findings underscore the promising potential of our approach to significantly enhance orthodontic diagnosis and treatment planning through automated landmark localization.

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

Enhanced Cephalometric Landmark Detection Using Multi-scale Feature Learning and Heatmap Regression

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Reeha Khan,
  • Muhammad Anwaar Khalid,
  • Kanwal Zulfiqar,
  • Ulfat Bashir,
  • Muhammad Moazam Fraz

摘要

In orthodontic diagnosis and treatment planning, accurate localization of landmarks in X-ray cephalograms is essential. While deep learning approaches have shown promise in automating landmark identification, ensuring consistent performance and strong generalization remains challenging due to the lack of diverse datasets. This study presents a novel architecture based on Multi-Scale Convolutional Neural Network (CNN) and Heatmap Regression for automatic cephalometric landmark detection. This method captures global relationships among landmarks as well as the distinct characteristics of individual landmarks, allowing for discriminative analysis across multiple scales. We utilize a newly developed benchmark dataset named “Aariz”. The proposed framework attained an overall mean radial error (MRE) of 1.942 ± 3.782 on the test dataset, with the success detection rates 69.01% and 91.33% within the 2 mm and 4 mm ranges, respectively. Additionally, we compared the performance of coordinate regression and heat map regression. These findings underscore the promising potential of our approach to significantly enhance orthodontic diagnosis and treatment planning through automated landmark localization.