Scholarlink

<p>Breast cancer is one among the prominent origins of mortality in females aged 20 to 59. Early detection significantly improves survival rates, with an 80% chance in early stages, dropping to around 40% in later stages. Hematoxylin and Eosin stained histopathology images are commonly used for diagnosis, but manual analysis is time-consuming and prone to variability. Automated prediction with deep learning, predominantly Convolutional Neural Networks (CNNs), has gained attention for improving diagnostic accuracy. However, training CNNs from scratch necessitates large labeled datasets, which are often unavailable in medical imaging, and demand high computational resources. Transfer learning, using pre-trained models like ResNet50, GoogleNet, AlexNet, and DenseNet from natural image datasets like ImageNet, offers a solution. In this approach, three architectures namely AlexNet, GoogleNet, and DenseNet121 are adjusted for breast cancer image prediction. Custom dense layers are added, and the last two layers are fine-tuned using the BreakHis and MIAS dataset. Experimental results demonstrate that transfer learning outperforms models trained from scratch, with DenseNet121 achieving the highest accuracy of 96.56%. This highlights the prospective of transfer learning in refining breast cancer detection from histopathological images. This study uniquely incorporates ROC–AUC and inference-time analysis, ensuring both diagnostic reliability and clinical applicability.</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

Leveraging transfer learning with convolutional neural networks for enhanced breast cancer detection

  • 期刊论文

  • 发表日期：

    2026年4月15日

查看全文
  • K Ramalakshmi,
  • Vinod Kumar Shukla,
  • L Krishna Kumari

摘要

Breast cancer is one among the prominent origins of mortality in females aged 20 to 59. Early detection significantly improves survival rates, with an 80% chance in early stages, dropping to around 40% in later stages. Hematoxylin and Eosin stained histopathology images are commonly used for diagnosis, but manual analysis is time-consuming and prone to variability. Automated prediction with deep learning, predominantly Convolutional Neural Networks (CNNs), has gained attention for improving diagnostic accuracy. However, training CNNs from scratch necessitates large labeled datasets, which are often unavailable in medical imaging, and demand high computational resources. Transfer learning, using pre-trained models like ResNet50, GoogleNet, AlexNet, and DenseNet from natural image datasets like ImageNet, offers a solution. In this approach, three architectures namely AlexNet, GoogleNet, and DenseNet121 are adjusted for breast cancer image prediction. Custom dense layers are added, and the last two layers are fine-tuned using the BreakHis and MIAS dataset. Experimental results demonstrate that transfer learning outperforms models trained from scratch, with DenseNet121 achieving the highest accuracy of 96.56%. This highlights the prospective of transfer learning in refining breast cancer detection from histopathological images. This study uniquely incorporates ROC–AUC and inference-time analysis, ensuring both diagnostic reliability and clinical applicability.