Scholarlink

<p>Cancer remains a leading global health challenge, necessitating early and accurate detection of cancerous cells to improve patient outcomes. This study proposes a Transfer Learning (TL) approach to classify viable (live) and non-viable (dead) cells using pre-trained Convolutional Neural Network (CNN) models. Leveraging TL enables efficient training by adapting knowledge from large-scale datasets, reducing the need for extensive labeled data and computational resources. We evaluate six state-of-the-art TL models on a binary dataset containing over 21,000 labeled cell (JIMT breast cancer cells) images. Among them, an optimized EfficientNetB0 model achieves the highest accuracy, between 87.71% and 89.90%, demonstrating robust performance with minimal misclassifications. The optimized EfficientNetB0 model achieved a mean accuracy of approximately 89.90% with an estimated standard deviation of ± 0.73% after 20–40 runs, indicating stable performance across runs while it maintains 89.92 of precision, 89.83 of recall, and 89.87 of f1-score. VGG16 ranks as the second-best but shows variable consistency.</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

AI-powered cancer diagnosis: classifying viable (live) vs non-viable (dead) cells using transfer learning

  • 期刊论文

  • 发表日期：

    2025年10月28日

查看全文
  • Saif Ur Rehman Khan,
  • Hafeez Ur Rehman,
  • Omair Bilal

摘要

Cancer remains a leading global health challenge, necessitating early and accurate detection of cancerous cells to improve patient outcomes. This study proposes a Transfer Learning (TL) approach to classify viable (live) and non-viable (dead) cells using pre-trained Convolutional Neural Network (CNN) models. Leveraging TL enables efficient training by adapting knowledge from large-scale datasets, reducing the need for extensive labeled data and computational resources. We evaluate six state-of-the-art TL models on a binary dataset containing over 21,000 labeled cell (JIMT breast cancer cells) images. Among them, an optimized EfficientNetB0 model achieves the highest accuracy, between 87.71% and 89.90%, demonstrating robust performance with minimal misclassifications. The optimized EfficientNetB0 model achieved a mean accuracy of approximately 89.90% with an estimated standard deviation of ± 0.73% after 20–40 runs, indicating stable performance across runs while it maintains 89.92 of precision, 89.83 of recall, and 89.87 of f1-score. VGG16 ranks as the second-best but shows variable consistency.