Scholarlink

<p>Accurate and timely diagnosis of brain tumors is crucial for effective treatment planning, yet the rapid growth of medical imaging data presents significant challenges for manual interpretation. This paper introduces a novel, scalable hybrid deep learning framework designed for robust brain tumor classification from large-scale MRI datasets. Our architecture distinctively combines a Convolutional Neural Network (CNN) for fine-grained local feature extraction with a Vision Transformer (ViT) for modeling global contextual dependencies. To specifically address the challenge of tumor heterogeneity, we introduce a Mixture-of-Experts (MoE) layer that dynamically routes features to specialized subnetworks. Feature representation is further refined by a Multi-Head Latent Attention (MHLA) mechanism, which focuses on the most salient diagnostic information. A key aspect of our methodology is an iterative data-centric refinement strategy to enhance label reliability and reduce intra-class variability in large, potentially noisy datasets. Evaluated on a comprehensive four-class benchmark dataset (glioma, meningioma, pituitary, and no tumor), our model achieves state-of-the-art performance with 98.9% accuracy. This work contributes a scalable, accurate, and more interpretable AI pipeline that directly addresses the complexities of real-world medical big data, demonstrating a clear novelty in its synergistic integration of advanced neural components for a challenging clinical problem.</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

A Scalable Brain Tumor Diagnosis from Large-Scale MRI Datasets Using CNN-ViT and Expert-Attention Fusions

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月7日

查看全文
  • Ram Kumar Yadav,
  • Manoj Kumar,
  • Avishek Nandi

摘要

Accurate and timely diagnosis of brain tumors is crucial for effective treatment planning, yet the rapid growth of medical imaging data presents significant challenges for manual interpretation. This paper introduces a novel, scalable hybrid deep learning framework designed for robust brain tumor classification from large-scale MRI datasets. Our architecture distinctively combines a Convolutional Neural Network (CNN) for fine-grained local feature extraction with a Vision Transformer (ViT) for modeling global contextual dependencies. To specifically address the challenge of tumor heterogeneity, we introduce a Mixture-of-Experts (MoE) layer that dynamically routes features to specialized subnetworks. Feature representation is further refined by a Multi-Head Latent Attention (MHLA) mechanism, which focuses on the most salient diagnostic information. A key aspect of our methodology is an iterative data-centric refinement strategy to enhance label reliability and reduce intra-class variability in large, potentially noisy datasets. Evaluated on a comprehensive four-class benchmark dataset (glioma, meningioma, pituitary, and no tumor), our model achieves state-of-the-art performance with 98.9% accuracy. This work contributes a scalable, accurate, and more interpretable AI pipeline that directly addresses the complexities of real-world medical big data, demonstrating a clear novelty in its synergistic integration of advanced neural components for a challenging clinical problem.