Scholarlink

Background <p>Oral squamous cell carcinoma (OSCC) is a major health concern due to delayed diagnosis. Deep learning (DL) offers automated detection and classification of oral lesions, including oral potentially malignant disorders (OPMDs).</p> Methods <p>A structured narrative review was conducted by searching PubMed, Web of Science, Scopus, and Google Scholar for peer-reviewed studies published between January 2020 and October 2025. Fifteen eligible studies using deep learning on clinical or smartphone images were qualitatively synthesized.</p> Results <p>Sensitivity ranged 55–98%, specificity 80-96.5%, with AUCs above 0.86. CNN models reached 80–95% accuracy, and some hybrid models exceeded 95%. Segmentation and object detection models achieved Dice coefficients up to 0.95 and mean average precision values of 74–85%.</p> Conclusion <p>DL shows promise for early, non-invasive oral cancer and OPMD diagnosis. Limitations include small, imbalanced datasets and limited external validation. Unlike broader AI reviews, this review specifically compares recent image-based DL architectures and discusses their implications for clinical translation. Future research should prioritize multicenter datasets, external validation, explainable AI, and prospective clinical evaluation.</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

Artificial intelligence and deep learning applications in oral oncology

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月31日

查看全文
  • Mina Khayamzadeh,
  • Yasmina Sakhipour,
  • Dorsa Samani

摘要

Background

Oral squamous cell carcinoma (OSCC) is a major health concern due to delayed diagnosis. Deep learning (DL) offers automated detection and classification of oral lesions, including oral potentially malignant disorders (OPMDs).

Methods

A structured narrative review was conducted by searching PubMed, Web of Science, Scopus, and Google Scholar for peer-reviewed studies published between January 2020 and October 2025. Fifteen eligible studies using deep learning on clinical or smartphone images were qualitatively synthesized.

Results

Sensitivity ranged 55–98%, specificity 80-96.5%, with AUCs above 0.86. CNN models reached 80–95% accuracy, and some hybrid models exceeded 95%. Segmentation and object detection models achieved Dice coefficients up to 0.95 and mean average precision values of 74–85%.

Conclusion

DL shows promise for early, non-invasive oral cancer and OPMD diagnosis. Limitations include small, imbalanced datasets and limited external validation. Unlike broader AI reviews, this review specifically compares recent image-based DL architectures and discusses their implications for clinical translation. Future research should prioritize multicenter datasets, external validation, explainable AI, and prospective clinical evaluation.