Scholarlink

Oral cancer remains a significant global health challenge, with the highest prevalence observed in low and middle-income countries. Early detection and diagnosis are critical for improving outcomes, yet current methods often lack accessibility and affordability in resource-constrained settings. To address this, we developed a deep learning-based framework to classify white light oral cavity images as Suspicious or Non-Suspicious. Utilizing a dataset representative of the Sri Lankan population, we evaluated the performance of several convolutional neural networks, leveraging transfer learning for enhanced feature extraction and classification. Our proposed architecture, integrating a novel voting-based algorithm, outperformed existing methods. Additionally, Monte Carlo Dropout was employed to quantify the uncertainty in model predictions, ensuring robust performance assessment. The model achieved a recall of 0.8650, an F1-score of 0.8180, and an accuracy of 80.75% on the test data, demonstrating its effectiveness. This work highlights the potential of deep learning to enable low-cost, automated, and accurate early detection of oral cancer, particularly in underserved regions.

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

Automated Screening of Oral Cancer: Classifying Suspicious Versus Non-suspicious White Light Images

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Astha Patidar,
  • Suparna Podder,
  • M. Vani,
  • Jeny Rajan

摘要

Oral cancer remains a significant global health challenge, with the highest prevalence observed in low and middle-income countries. Early detection and diagnosis are critical for improving outcomes, yet current methods often lack accessibility and affordability in resource-constrained settings. To address this, we developed a deep learning-based framework to classify white light oral cavity images as Suspicious or Non-Suspicious. Utilizing a dataset representative of the Sri Lankan population, we evaluated the performance of several convolutional neural networks, leveraging transfer learning for enhanced feature extraction and classification. Our proposed architecture, integrating a novel voting-based algorithm, outperformed existing methods. Additionally, Monte Carlo Dropout was employed to quantify the uncertainty in model predictions, ensuring robust performance assessment. The model achieved a recall of 0.8650, an F1-score of 0.8180, and an accuracy of 80.75% on the test data, demonstrating its effectiveness. This work highlights the potential of deep learning to enable low-cost, automated, and accurate early detection of oral cancer, particularly in underserved regions.