Scholarlink

<p>Brain and central nervous system cancers have high mortality rates, making early detection essential for improving patient outcomes. This study employs two convolutional neural network (CNN) models, AlexNet and ResNet-18, to classify brain MRI images into four categories: Pituitary tumors, Meningiomas, Gliomas, and No Tumor. To enhance diagnostic accuracy and reduce experimental time, the Taguchi method is applied to optimize key CNN parameters, including learning rate, optimizer, pooling method, convolution kernel size, and mini-batch size. Experimental results demonstrate that parameter optimization using the Taguchi method significantly improves classification performance. ResNet-18 achieved an accuracy of 92.19% with a training time of 3&#xa0;min and 31&#xa0;s, comparable to ResNet-50 but with significantly lower computational cost. After optimization, ResNet-18’s accuracy increased to 95.31%, with a reduced training time of 3&#xa0;min and 29&#xa0;s, making it the best-performing model in this study.</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

Enhancing diagnostic accuracy of brain and CNS tumors through CNN parameter optimization using Taguchi’s method

  • 期刊论文

  • 发表日期：

    2025年8月1日

查看全文
  • Yu-Chi Li,
  • Chien-Wei Wu,
  • Fatt-Yang Chew

摘要

Brain and central nervous system cancers have high mortality rates, making early detection essential for improving patient outcomes. This study employs two convolutional neural network (CNN) models, AlexNet and ResNet-18, to classify brain MRI images into four categories: Pituitary tumors, Meningiomas, Gliomas, and No Tumor. To enhance diagnostic accuracy and reduce experimental time, the Taguchi method is applied to optimize key CNN parameters, including learning rate, optimizer, pooling method, convolution kernel size, and mini-batch size. Experimental results demonstrate that parameter optimization using the Taguchi method significantly improves classification performance. ResNet-18 achieved an accuracy of 92.19% with a training time of 3 min and 31 s, comparable to ResNet-50 but with significantly lower computational cost. After optimization, ResNet-18’s accuracy increased to 95.31%, with a reduced training time of 3 min and 29 s, making it the best-performing model in this study.