Scholarlink

<p>Brain tumors, significant and often fatal, have substantial socio-economic effects. Distinguishing between various types, such as Gliomas, Meningiomas, and Pituitary Tumors, using MRI data is vital for aiding radiologists and preventing unnecessary, risky biopsies. This study aims to develop an automated system using the Vision Transformer (ViT) to predict and classify six brain tumor types: Glioma, Meningioma, Neurocitoma, Normal, Outros Tipos De Lesoes, and Schwannoma. The ViT model captures global information by processing image patches in parallel, reducing computational complexity. The model’s 12 encoders, with self-attention mechanisms, create attention maps to highlight critical features, enhancing MRI image classification. The network is trained with various optimizers—Adam, RMSprop, and SGDM—using different learning rates and hyperparameters. The features identified by the ViT network were validated using the LIME approach, showing superior performance to traditional methods with an accuracy of 97.4% at an optimal learning rate of 0.0001.</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

Optimizing Brain Tumor Classification with ViT: A Metaheuristic Approach

  • 期刊论文

  • 发表日期：

    2025年6月30日

查看全文
  • Vishnupriya S,
  • Sakthisudhan Karuppanan

摘要

Brain tumors, significant and often fatal, have substantial socio-economic effects. Distinguishing between various types, such as Gliomas, Meningiomas, and Pituitary Tumors, using MRI data is vital for aiding radiologists and preventing unnecessary, risky biopsies. This study aims to develop an automated system using the Vision Transformer (ViT) to predict and classify six brain tumor types: Glioma, Meningioma, Neurocitoma, Normal, Outros Tipos De Lesoes, and Schwannoma. The ViT model captures global information by processing image patches in parallel, reducing computational complexity. The model’s 12 encoders, with self-attention mechanisms, create attention maps to highlight critical features, enhancing MRI image classification. The network is trained with various optimizers—Adam, RMSprop, and SGDM—using different learning rates and hyperparameters. The features identified by the ViT network were validated using the LIME approach, showing superior performance to traditional methods with an accuracy of 97.4% at an optimal learning rate of 0.0001.