Scholarlink

<p>Breast cancer becomes a challenging issue and most critical health concerns among women around the world. Timely diagnosis, especially using mammography, increases the probability of effective treatment and long-time survival. The need for automated tools using modern deep-learning techniques comes into the picture because examining mammogram images can sometimes lead to errors. This work focuses on describing how breast-cancer classification can be improved by using widely used CNN architectures through transfer learning. As we are using medical images, despite using these networks in their standard form because they are typically designed for broad image analysis tasks, we modified models—EfficientNet-B0, ResNet-50, and VGG-16—so they could better handle and capture the characteristics and give good results. With different evaluation metrics, these CNN models are examined to see how they boost the diagnostic accuracy. Before training, we resize all the images to a particular scale and CLAHE is used for contrast enhancement for making subtle details more visible. The modifications which we have made to the classifier heads are an important part to enhance the model’s potential for prediction. Among all three tested models, the highest accuracy achieved by EfficientNet-B0 is 97.77%. The ablation study also performed which shows its robustness when compared with the other architectures. The study explains why such architectural modifications are needed and how they are contributing to build more dependable automated diagnostic systems for breast cancer screening in actual healthcare environments.</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

Transfer learning driven breast cancer classification with modified CNN architectures using mammographic image data

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月27日

查看全文
  • Pragya Singh,
  • Sanjeev Kumar,
  • Yadvendra Pratap Singh

摘要

Breast cancer becomes a challenging issue and most critical health concerns among women around the world. Timely diagnosis, especially using mammography, increases the probability of effective treatment and long-time survival. The need for automated tools using modern deep-learning techniques comes into the picture because examining mammogram images can sometimes lead to errors. This work focuses on describing how breast-cancer classification can be improved by using widely used CNN architectures through transfer learning. As we are using medical images, despite using these networks in their standard form because they are typically designed for broad image analysis tasks, we modified models—EfficientNet-B0, ResNet-50, and VGG-16—so they could better handle and capture the characteristics and give good results. With different evaluation metrics, these CNN models are examined to see how they boost the diagnostic accuracy. Before training, we resize all the images to a particular scale and CLAHE is used for contrast enhancement for making subtle details more visible. The modifications which we have made to the classifier heads are an important part to enhance the model’s potential for prediction. Among all three tested models, the highest accuracy achieved by EfficientNet-B0 is 97.77%. The ablation study also performed which shows its robustness when compared with the other architectures. The study explains why such architectural modifications are needed and how they are contributing to build more dependable automated diagnostic systems for breast cancer screening in actual healthcare environments.