Scholarlink

<p>Oral potentially malignant disorders (OPMDs), particularly leukoplakia, represent critical precursors to oral cancer requiring systematic monitoring. However, current diagnostic approaches rely on invasive biopsies and subjective clinical assessment, which limit screening accessibility. Here, we report a deep learning system that provides both highly reliable diagnostic classification and visual explainability for OPMD assessment. Trained on 778 expertly annotated clinical images, our framework first utilizes a diagnostic classifier to distinguish normal mucosa, leukoplakia, and malignant transformation (accuracy: 91.1%, 95% CI: 86.7−94.2%). For cases flagged as abnormal, it conditionally triggers an interpretable segmentation module providing pixel-level lesion delineation (mean average precision: 72.2%). External validation on 193 independent clinical cases demonstrated maintained diagnostic accuracy (90.5%, 95% CI: 85.7−94.0%), with the strict lower confidence bounds significantly outperforming state-of-the-art general vision baselines. The system processes images in 3-5 milliseconds on standard GPUs (and approximately 1 second on conventional CPUs), enabling highly flexible point-of-care deployment. By combining diagnostic accuracy with dual-dimensional (spatial and mechanistic) explainability, this framework addresses key barriers to clinical AI adoption, offering a scalable and practical tool for OPMD screening and longitudinal monitoring in diverse healthcare settings.</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

An Interpretable AI System for Oral Leukoplakia Progression: From Early Screening to Lesion Delineation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月7日

查看全文
  • Linfei Feng,
  • Guanyu Chen,
  • Huabao Chen,
  • Susu Luo,
  • Xuanyu Li,
  • Aokun Liu,
  • Jinyue Li,
  • Huarui Liu,
  • Yimou Wang,
  • Feng He,
  • Lin Jiao,
  • Yang Liu,
  • Yani Zhang,
  • Qiankun Li

摘要

Oral potentially malignant disorders (OPMDs), particularly leukoplakia, represent critical precursors to oral cancer requiring systematic monitoring. However, current diagnostic approaches rely on invasive biopsies and subjective clinical assessment, which limit screening accessibility. Here, we report a deep learning system that provides both highly reliable diagnostic classification and visual explainability for OPMD assessment. Trained on 778 expertly annotated clinical images, our framework first utilizes a diagnostic classifier to distinguish normal mucosa, leukoplakia, and malignant transformation (accuracy: 91.1%, 95% CI: 86.7−94.2%). For cases flagged as abnormal, it conditionally triggers an interpretable segmentation module providing pixel-level lesion delineation (mean average precision: 72.2%). External validation on 193 independent clinical cases demonstrated maintained diagnostic accuracy (90.5%, 95% CI: 85.7−94.0%), with the strict lower confidence bounds significantly outperforming state-of-the-art general vision baselines. The system processes images in 3-5 milliseconds on standard GPUs (and approximately 1 second on conventional CPUs), enabling highly flexible point-of-care deployment. By combining diagnostic accuracy with dual-dimensional (spatial and mechanistic) explainability, this framework addresses key barriers to clinical AI adoption, offering a scalable and practical tool for OPMD screening and longitudinal monitoring in diverse healthcare settings.