Scholarlink

<p>Identification of malignant and non-malignant regions in breast cancer whole slide images (WSIs) is essential for understanding tumor heterogeneity and histopathological evaluation. Annotated data allows deep learning models to learn features and analyze histological structures. In this study, we obtained 50 breast cancer WSIs from The Cancer Genome Atlas (TCGA) and had an expert pathologist manually annotate malignant (<i>n</i> = 1,882) and non-malignant (<i>n</i> = 374) regions using QuPath. The annotated regions were independently reviewed by a second pathologist, achieving an inter-observer agreement of 99.95%. In addition, to assess annotations’ quality, we trained a hybrid contrastive-supervised machine learning pipeline for patch-level malignant vs. non-malignant classification. The model achieved a high F1-score of 0.90 indicating that our annotations are comparable in quality to those presented in public datasets. The proposed dataset provides expert-quality annotations and a unique resource for benchmarking and developing AI models for breast cancer histopathology.</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

Malignant vs. Non-malignant Annotations on TCGA Breast Cancer Whole Slide Images for AI Analysis

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月2日

查看全文
  • Zahra Alidousti Shahraki,
  • Otto Jokelainen,
  • Mari Valkonen,
  • Päivi Auvinen,
  • Arto Mannermaa,
  • Hamid Behravan

摘要

Identification of malignant and non-malignant regions in breast cancer whole slide images (WSIs) is essential for understanding tumor heterogeneity and histopathological evaluation. Annotated data allows deep learning models to learn features and analyze histological structures. In this study, we obtained 50 breast cancer WSIs from The Cancer Genome Atlas (TCGA) and had an expert pathologist manually annotate malignant (n = 1,882) and non-malignant (n = 374) regions using QuPath. The annotated regions were independently reviewed by a second pathologist, achieving an inter-observer agreement of 99.95%. In addition, to assess annotations’ quality, we trained a hybrid contrastive-supervised machine learning pipeline for patch-level malignant vs. non-malignant classification. The model achieved a high F1-score of 0.90 indicating that our annotations are comparable in quality to those presented in public datasets. The proposed dataset provides expert-quality annotations and a unique resource for benchmarking and developing AI models for breast cancer histopathology.