Scholarlink

<p>Transformer-based deep learning models have garnered significant attention in recent years for cancer medical image classification. As cancer incidence rates continue to rise globally, the development of accurate and efficient diagnostic imaging methods has become increasingly critical. Despite their promise, existing Transformer architectures encounter several limitations in medical image classification applications, including high computational complexity, suboptimal integration of local and global features, and insufficient spatial prior knowledge incorporation. To address these limitations, a Multi-Channel Feature Fusion Attention (MCFA) mechanism is proposed to be embedded into the Retentive Networks Meet Vision Transformers (RMT) architecture. This new architecture employs Haar wavelet downsampling (HWD) to enhance local feature representation and expands receptive field coverage through 7 × 7 convolution kernels in Convolutional Neural Networks (CNN), effectively capturing rich contextual information while reducing feature map dimensions and retaining critical local information. Additionally, residual connections are integrated to mitigate gradient vanishing in deep networks. The key innovation is the cross-layer attention (CLAttn) mechanism. This mechanism integrates different feature maps through grouped interactions and employs linear-complexity self-attention to achieve deep fusion of features from CNN and Haar wavelet downsampling. Consequently, the model demonstrates significantly improved feature recognition capabilities and classification performance when processing heterogeneous regions in colon tissue samples. Experimental evaluations reveal that the proposed MCFA RMT (MFRMT) model outperforms state-of-the-art Transformer architectures across key metrics, including accuracy (Acc1) and quadratic-weighted kappa (Q-kappa).</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

MFRMT: feature fusion meet transformer for cancer classification

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • Pengfeng Li,
  • Qiang Wang,
  • Shulin Cui,
  • Yang Yin,
  • Yifan Wang,
  • Weibo Song

摘要

Transformer-based deep learning models have garnered significant attention in recent years for cancer medical image classification. As cancer incidence rates continue to rise globally, the development of accurate and efficient diagnostic imaging methods has become increasingly critical. Despite their promise, existing Transformer architectures encounter several limitations in medical image classification applications, including high computational complexity, suboptimal integration of local and global features, and insufficient spatial prior knowledge incorporation. To address these limitations, a Multi-Channel Feature Fusion Attention (MCFA) mechanism is proposed to be embedded into the Retentive Networks Meet Vision Transformers (RMT) architecture. This new architecture employs Haar wavelet downsampling (HWD) to enhance local feature representation and expands receptive field coverage through 7 × 7 convolution kernels in Convolutional Neural Networks (CNN), effectively capturing rich contextual information while reducing feature map dimensions and retaining critical local information. Additionally, residual connections are integrated to mitigate gradient vanishing in deep networks. The key innovation is the cross-layer attention (CLAttn) mechanism. This mechanism integrates different feature maps through grouped interactions and employs linear-complexity self-attention to achieve deep fusion of features from CNN and Haar wavelet downsampling. Consequently, the model demonstrates significantly improved feature recognition capabilities and classification performance when processing heterogeneous regions in colon tissue samples. Experimental evaluations reveal that the proposed MCFA RMT (MFRMT) model outperforms state-of-the-art Transformer architectures across key metrics, including accuracy (Acc1) and quadratic-weighted kappa (Q-kappa).