Scholarlink

The early and precise detection of brain tumors from Magnetic Resonance Imaging (MRI) is a crucial component of diagnostic medicine and treatment strategy [5]. Deep learning, particularly Convolutional Neural Networks (CNNs), offers a powerful means of automating this analysis, potentially increasing both the speed and consistency of detection [9]. This paper conducts a comparative evaluation of three CNN architectures—a custom-designed CNN, ResNet-50, and EfficientNet-B0—on publicly available MRI datasets for both binary (tumor vs. no tumor) and multi-class (glioma, meningioma, pituitary, no tumor) classification tasks. All MRI scans were standardized to 224 × 224 pixels, normalized, and augmented with random flips and rotations to improve generalization [2]. A comprehensive grid search optimized hyperparameters, including learning rate, optimizer, and dropout rate. Results indicate that for binary classification, the pre-trained ResNet-50 model achieved superior performance with an accuracy of 94.7% and an AUC of 0.984. Conversely, for the more complex multi-class task, the optimized custom CNN (∼1.5 million parameters) outperformed the larger models, achieving 90.2% accuracy. These findings suggest that while large, pre-trained models excel at simpler detection tasks, smaller, well-regularized custom CNNs can offer better generalization on complex, multi-class problems with limited data.

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

An Experimental Study of Deep Learning Models for Brain Tumor Classification in MRI

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Aarti Jain,
  • Nikita Bhatt

摘要

The early and precise detection of brain tumors from Magnetic Resonance Imaging (MRI) is a crucial component of diagnostic medicine and treatment strategy [5]. Deep learning, particularly Convolutional Neural Networks (CNNs), offers a powerful means of automating this analysis, potentially increasing both the speed and consistency of detection [9]. This paper conducts a comparative evaluation of three CNN architectures—a custom-designed CNN, ResNet-50, and EfficientNet-B0—on publicly available MRI datasets for both binary (tumor vs. no tumor) and multi-class (glioma, meningioma, pituitary, no tumor) classification tasks. All MRI scans were standardized to 224 × 224 pixels, normalized, and augmented with random flips and rotations to improve generalization [2]. A comprehensive grid search optimized hyperparameters, including learning rate, optimizer, and dropout rate. Results indicate that for binary classification, the pre-trained ResNet-50 model achieved superior performance with an accuracy of 94.7% and an AUC of 0.984. Conversely, for the more complex multi-class task, the optimized custom CNN (∼1.5 million parameters) outperformed the larger models, achieving 90.2% accuracy. These findings suggest that while large, pre-trained models excel at simpler detection tasks, smaller, well-regularized custom CNNs can offer better generalization on complex, multi-class problems with limited data.