Scholarlink

Background and objective <p>Intraoral photographs are routinely taken in orthodontic practice and provide valuable visual information for diagnostic purposes. To support the training of dental graduate students and prospective orthodontists in diagnosing sagittal malocclusions, this study aimed to develop an artificial intelligence (AI) model to classify sagittal dental malocclusions from digital intraoral photographs using the widely accepted Angle classification system, and to evaluate the explainability of the model’s decisions using three explainable AI (XAI) methods.</p> Materials and methods <p>A total of 5266 clinical RGB images showing dental occlusion from the lateral view of first-time orthodontic patients were retrieved from the clinic’s image database and classified by orthodontic experts into Angle Class I (2322 images; 44%), Class II (1880 images; 36%), and Class III (1064 images; 20%).The dataset was then divided into a training set of 4280 images. The validation set contained 474 images, and the test set comprised 512 images of a deep-learning classification model (VGG-11). The employed deep-learning classification model (VGG-11) was then trained and evaluated. Three XAI methods (Layerwise Relevance Propagation, PatternNet, and PatternAttribution) were used to generate heatmaps highlighting relevant areas for classification.</p> Results <p>The deep learning model correctly classified 75% of the test set images, achieving high predictive performance across all three Angle classes, with the area-under-the-curve being 0.91, 0.90, and 0.91 for Angle classes I, II, and III, respectively, indicating substantial discriminative ability. The most frequent misclassifications were Angle Class I being misclassified as Angle Class III, and Angle Classes II and III as Angle Class I. XAI highlighted the area surrounding the first molars as decisive for classification, although the three different XAI methods utilized different areas.</p> Conclusion <p>Deep learning proved effective for classifying dental malocclusion into Angle classes I, II and III using intraoral photographs. XAI revealed that the classification was based on clinically relevant features. Different XAI methods reflected on different features; combining more XAI methods may allow comprehensive assessment of a model’s classification logic and may accelerate the transfer into clinical application.</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

Deep learning for Angle classification based on intraoral photographs: an interpretability perspective

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月4日

查看全文
  • Petra Julia Koch,
  • José Eduardo Cejudo Grano de Oro,
  • Martha Büttner,
  • Lubaina Tayeb Arsiwala-Scheppach,
  • Julia De Geer,
  • Henrik Meyer-Lueckel,
  • Falk Schwendicke

摘要

Background and objective

Intraoral photographs are routinely taken in orthodontic practice and provide valuable visual information for diagnostic purposes. To support the training of dental graduate students and prospective orthodontists in diagnosing sagittal malocclusions, this study aimed to develop an artificial intelligence (AI) model to classify sagittal dental malocclusions from digital intraoral photographs using the widely accepted Angle classification system, and to evaluate the explainability of the model’s decisions using three explainable AI (XAI) methods.

Materials and methods

A total of 5266 clinical RGB images showing dental occlusion from the lateral view of first-time orthodontic patients were retrieved from the clinic’s image database and classified by orthodontic experts into Angle Class I (2322 images; 44%), Class II (1880 images; 36%), and Class III (1064 images; 20%).The dataset was then divided into a training set of 4280 images. The validation set contained 474 images, and the test set comprised 512 images of a deep-learning classification model (VGG-11). The employed deep-learning classification model (VGG-11) was then trained and evaluated. Three XAI methods (Layerwise Relevance Propagation, PatternNet, and PatternAttribution) were used to generate heatmaps highlighting relevant areas for classification.

Results

The deep learning model correctly classified 75% of the test set images, achieving high predictive performance across all three Angle classes, with the area-under-the-curve being 0.91, 0.90, and 0.91 for Angle classes I, II, and III, respectively, indicating substantial discriminative ability. The most frequent misclassifications were Angle Class I being misclassified as Angle Class III, and Angle Classes II and III as Angle Class I. XAI highlighted the area surrounding the first molars as decisive for classification, although the three different XAI methods utilized different areas.

Conclusion

Deep learning proved effective for classifying dental malocclusion into Angle classes I, II and III using intraoral photographs. XAI revealed that the classification was based on clinically relevant features. Different XAI methods reflected on different features; combining more XAI methods may allow comprehensive assessment of a model’s classification logic and may accelerate the transfer into clinical application.