Scholarlink

Brain tumors are caused by abnormal cell development in the brain and can be both malignant and benign. Accurate diagnosis of brain tumor type is critical for treatment planning, but it is difficult due to the brain’s structural complexity. This study looks at fine-tuning pre-trained convolutional neural networks for brain tumor classification using magnetic resonance imaging (MRI) images. 7023 T1-weighted MRI images from Kaggle are used, with results indicating meningioma, glioma, pituitary tumor, or no malignancy. Four models named ResNet50V2, VGG19, InceptionV3, and Xception, are trained with ImageNet weights and fine-tuned on the MRI dataset. Data augmentation is used during training to achieve regularization. The convolutional basis is frozen, and new classification layers are learned to predict tumor kinds. Among the architectures, InceptionV3 achieves an impressive accuracy of \(99.28\%\) for the training set and \(98.17\%\) for the testing set in the epoch landscape, while VGG19 demonstrates strong performance with \(98\%\) accuracy in the classification report beating the other models. This highlights the viability of using deep transfer learning techniques to create automated tools that help radiologists analyze MRI data for abnormal findings connected to brain tumors. The work sheds light on the relative benefits of pre-trained models for fine-tuning medical imaging tasks.

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

Fine-Tuning Pre-trained Convolutional Neural Network Deep Transfer Learning for Brain Tumor Classification Using Magnetic Resonance Images

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Farzana Islam Jui,
  • Mohammod Abul Kashem,
  • Md. Monirul Islam,
  • Jia Uddin

摘要

Brain tumors are caused by abnormal cell development in the brain and can be both malignant and benign. Accurate diagnosis of brain tumor type is critical for treatment planning, but it is difficult due to the brain’s structural complexity. This study looks at fine-tuning pre-trained convolutional neural networks for brain tumor classification using magnetic resonance imaging (MRI) images. 7023 T1-weighted MRI images from Kaggle are used, with results indicating meningioma, glioma, pituitary tumor, or no malignancy. Four models named ResNet50V2, VGG19, InceptionV3, and Xception, are trained with ImageNet weights and fine-tuned on the MRI dataset. Data augmentation is used during training to achieve regularization. The convolutional basis is frozen, and new classification layers are learned to predict tumor kinds. Among the architectures, InceptionV3 achieves an impressive accuracy of \(99.28\%\) for the training set and \(98.17\%\) for the testing set in the epoch landscape, while VGG19 demonstrates strong performance with \(98\%\) accuracy in the classification report beating the other models. This highlights the viability of using deep transfer learning techniques to create automated tools that help radiologists analyze MRI data for abnormal findings connected to brain tumors. The work sheds light on the relative benefits of pre-trained models for fine-tuning medical imaging tasks.