Scholarlink

<p>Alzheimer’s disease (AD) is a progressive neurodegenerative disorder characterized by cognitive decline, with early detection remaining a significant challenge due to subtle structural alterations in brain MRI scans. To address this, we propose a novel MRI classification model that effectively integrates local morphological features with global contextual information, enhancing diagnostic performance. Our method incorporates multi-scale convolutional layers to capture fine-grained local patterns, dilated convolutions to expand receptive fields for enriched global context, and an advanced attention mechanism combining channel attention with an improved SimAM module to facilitate intelligent feature fusion. This synergistic design enables a more comprehensive representation of both localized structural variations and their broader implications, which is crucial for early-stage detection. Furthermore, our framework employs three specialized binary classifiers, each leveraging the proposed fusion strategy to enhance classification robustness. Experimental results demonstrate significant performance gains, achieving 86.7% accuracy for AD, 92.6% for mild cognitive impairment (MCI), and 86.4% for normal control (NC) classification. Notably, our approach outperforms existing methods, particularly in distinguishing early-stage abnormalities where the interplay between local and global features is most critical.</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

Local–global feature fusion for enhanced Alzheimer’s detection: a multi-scale attention-driven MRI classification model

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年7月22日

查看全文
  • Kerang Cao,
  • Zhongqing Lu,
  • Kaidi Zhang,
  • Yu Liu,
  • Ye Ma,
  • Jun Wang,
  • Hoekyung Jung

摘要

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder characterized by cognitive decline, with early detection remaining a significant challenge due to subtle structural alterations in brain MRI scans. To address this, we propose a novel MRI classification model that effectively integrates local morphological features with global contextual information, enhancing diagnostic performance. Our method incorporates multi-scale convolutional layers to capture fine-grained local patterns, dilated convolutions to expand receptive fields for enriched global context, and an advanced attention mechanism combining channel attention with an improved SimAM module to facilitate intelligent feature fusion. This synergistic design enables a more comprehensive representation of both localized structural variations and their broader implications, which is crucial for early-stage detection. Furthermore, our framework employs three specialized binary classifiers, each leveraging the proposed fusion strategy to enhance classification robustness. Experimental results demonstrate significant performance gains, achieving 86.7% accuracy for AD, 92.6% for mild cognitive impairment (MCI), and 86.4% for normal control (NC) classification. Notably, our approach outperforms existing methods, particularly in distinguishing early-stage abnormalities where the interplay between local and global features is most critical.