Scholarlink

Objectives <p>This study aimed to develop a hybrid deep learning model to automate the prognosis of impacted maxillary canines from panoramic radiographs. The model categorizes cases into three prognostic categories: good, average, and poor, as defined by expert orthodontists using clinical and radiographic criteria. Automating this process addresses the current clinical challenge of subjective assessment and variability among practitioners, thereby supporting decision-making for treatment planning.</p> Methods <p>The proposed architecture combines Swin Transformer and ResNet18 for feature extraction, integrated with an attention mechanism to enhance prognostic classification. This architecture was selected because Swin Transformer excels at capturing global contextual information, while ResNet18 effectively extracts local spatial features. A total of 294 panoramic radiographic images were used, each preprocessed using three image enhancement techniques: logarithmic transformation, Sobel edge detection, and median filtering.</p> Results <p>We compared the classification performance of the hybrid model with each image enhancement technique against that of the standalone Swin Transformer and ResNet18 models under the same conditions. The hybrid model using median filtering achieved the highest performance, with an AUC of 88.3%, outperforming both standalone architectures. Additionally, we conducted an ANOVA test to evaluate the statistical significance of differences in performance between deep learning models and image enhancement techniques. ANOVA results indicated that the hybrid model yielded a statistically significant improvement over individual models, although no significant differences were found when using image enhancement techniques. Additionally, the hybrid model surpassed the performance of three orthodontic residents in the same classification task.</p> Conclusions <p>The findings show that deep learning methods are effective for the prognostic assessment of maxillary canine impaction, providing a useful tool for dentists and orthodontists in clinical evaluation and decision-making.</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

Hybrid deep learning-based prognostic classification of maxillary canine impaction on panoramic radiographs

  • 期刊论文

  • 发表日期：

    2026年4月28日

查看全文
  • Roaa Alsubhi,
  • Hussam Alsharif,
  • Hasan Kadi,
  • Mohammed Barashi,
  • Hanadi Khalifa,
  • Reem Bashammakh

摘要

Objectives

This study aimed to develop a hybrid deep learning model to automate the prognosis of impacted maxillary canines from panoramic radiographs. The model categorizes cases into three prognostic categories: good, average, and poor, as defined by expert orthodontists using clinical and radiographic criteria. Automating this process addresses the current clinical challenge of subjective assessment and variability among practitioners, thereby supporting decision-making for treatment planning.

Methods

The proposed architecture combines Swin Transformer and ResNet18 for feature extraction, integrated with an attention mechanism to enhance prognostic classification. This architecture was selected because Swin Transformer excels at capturing global contextual information, while ResNet18 effectively extracts local spatial features. A total of 294 panoramic radiographic images were used, each preprocessed using three image enhancement techniques: logarithmic transformation, Sobel edge detection, and median filtering.

Results

We compared the classification performance of the hybrid model with each image enhancement technique against that of the standalone Swin Transformer and ResNet18 models under the same conditions. The hybrid model using median filtering achieved the highest performance, with an AUC of 88.3%, outperforming both standalone architectures. Additionally, we conducted an ANOVA test to evaluate the statistical significance of differences in performance between deep learning models and image enhancement techniques. ANOVA results indicated that the hybrid model yielded a statistically significant improvement over individual models, although no significant differences were found when using image enhancement techniques. Additionally, the hybrid model surpassed the performance of three orthodontic residents in the same classification task.

Conclusions

The findings show that deep learning methods are effective for the prognostic assessment of maxillary canine impaction, providing a useful tool for dentists and orthodontists in clinical evaluation and decision-making.