Scholarlink

<p>Deep learning architectures have revolutionized medical imaging through their capacity to automatically extract complex patterns. However, increasing the depth of these models often introduces challenges such as overfitting, vanishing gradients, and increased computational complexity. In order to resolve these concerns, the current study proposes MedLK-SCNet, a high-depth neural network designed for laryngeal cancer detection using multi-scale feature fusion and dual-pathway extraction. The architecture integrates a Simplified Large Kernel (Lark) Block and a Small Kernel (SmaK) Block to simultaneously capture global contextual relationships and local spatial details. Furthermore, the model integrates the ScK2 module, utilizing both Convolution in Spatial and Channel(ScConv), to reduce the computational demand while sustaining feature richness. Experimental evaluations on the laryngeal dataset demonstrate that the proposed architecture achieves robust generalization and superior diagnostic performance. The fusion of multi-scale feature vectors from the different stages ensures a comprehensive visual understanding, offering a significant advancement in automated larynx image classification.</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

MedLK-SCNet: a multi-scale hierarchical fusion model for high-precision classification of laryngeal malignancies

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月29日

查看全文
  • Puneet Misra,
  • Mohd Usman,
  • Siddharth Chaurasia,
  • Ginika Mahajan

摘要

Deep learning architectures have revolutionized medical imaging through their capacity to automatically extract complex patterns. However, increasing the depth of these models often introduces challenges such as overfitting, vanishing gradients, and increased computational complexity. In order to resolve these concerns, the current study proposes MedLK-SCNet, a high-depth neural network designed for laryngeal cancer detection using multi-scale feature fusion and dual-pathway extraction. The architecture integrates a Simplified Large Kernel (Lark) Block and a Small Kernel (SmaK) Block to simultaneously capture global contextual relationships and local spatial details. Furthermore, the model integrates the ScK2 module, utilizing both Convolution in Spatial and Channel(ScConv), to reduce the computational demand while sustaining feature richness. Experimental evaluations on the laryngeal dataset demonstrate that the proposed architecture achieves robust generalization and superior diagnostic performance. The fusion of multi-scale feature vectors from the different stages ensures a comprehensive visual understanding, offering a significant advancement in automated larynx image classification.