Scholarlink

<p>Oral squamous cell carcinoma (OSCC) is the most common oral malignancy, with high incidence in South and Southeast Asia. While histopathological examination of hematoxylin and eosin (H&amp;E) slides is the diagnostic gold standard, it is labor-intensive and prone to variability. This study explores YOLO-based deep learning models for automated binary segmentation of OSCC. We evaluate YOLOv11, YOLOv12 and Vanilla U-Net under baseline, manual, and Optuna-based fine-tuning, using both the public Oral Cavity-Derived Cancer (OCDC) dataset and an indigenous dataset from Northeast India, digitized with a cost-effective Micalys WSI system. YOLOv11 consistently outperformed YOLOv12, achieving Dice and IoU scores of 0.86 and 0.79 on OCDC and 0.57 and 0.51 on the indigenous dataset. Cross-dataset evaluation further confirmed clinical reliability, with predicted lesions aligning within annotated boundaries. These results highlight the potential of lightweight YOLO models for efficient, accurate, and regionally adaptable OSCC diagnosis support.</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

Automated Binary Segmentation of Oral Squamous Cell Carcinoma in H&E-Stained Histopathology Slides Using YOLO-Based Deep Learning Models

  • 期刊论文

  • 发表日期：

    2026年5月27日

查看全文
  • Rupesh Mandal,
  • Shubham Tarafder,
  • Nupur Choudhury,
  • Mrinmoy Mayur Choudhury,
  • Mukhtanjali Deka

摘要

Oral squamous cell carcinoma (OSCC) is the most common oral malignancy, with high incidence in South and Southeast Asia. While histopathological examination of hematoxylin and eosin (H&E) slides is the diagnostic gold standard, it is labor-intensive and prone to variability. This study explores YOLO-based deep learning models for automated binary segmentation of OSCC. We evaluate YOLOv11, YOLOv12 and Vanilla U-Net under baseline, manual, and Optuna-based fine-tuning, using both the public Oral Cavity-Derived Cancer (OCDC) dataset and an indigenous dataset from Northeast India, digitized with a cost-effective Micalys WSI system. YOLOv11 consistently outperformed YOLOv12, achieving Dice and IoU scores of 0.86 and 0.79 on OCDC and 0.57 and 0.51 on the indigenous dataset. Cross-dataset evaluation further confirmed clinical reliability, with predicted lesions aligning within annotated boundaries. These results highlight the potential of lightweight YOLO models for efficient, accurate, and regionally adaptable OSCC diagnosis support.