Scholarlink

<p>Developing computer-aided diagnosis systems that balance accuracy and efficiency remains a challenge in COVID-19 and respiratory disease classification. To develop effective lightweight models, we investigated various optimization strategies, including the convolutional block attention module, depthwise separable convolutions, global average pooling, residual connections, and knowledge distillation. Leveraging these optimization strategies, we introduce two models, LitCovNet 1 and LitCovNet 2, and evaluate them on the COVID-19 Radiography Database and COVID-QU-Ex datasets. LitCovNet 2 achieves 98.78% accuracy on the Radiography Database and 95.64% on COVID-QU-Ex with only 67,349 parameters, while LitCovNet 1 attains 95.70% on COVID-QU-Ex with 131,861 parameters. Despite being 350<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> </InlineEquation> smaller than ResNet50, LitCovNet 2 shows comparable accuracy and markedly superior FLOPs, model size, and inference speed. This makes it a perfect candidate for resource-constrained deployment. Benchmarking on multiple edge devices further demonstrates their suitability for resource-constrained deployment. These results establish LitCovNet as an effective lightweight solution for large-scale multiclass COVID-19 and respiratory disease 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

LitCovNet: Attention-Based Lightweight Convolutional Network for Covid and Lung Disease Classification from Chest X-ray Images

  • 期刊论文

  • 发表日期：

    2025年12月29日

查看全文
  • Md. Yearat Hossain,
  • Md. Mahbub Hasan Rakib,
  • Rashedur M. Rahman

摘要

Developing computer-aided diagnosis systems that balance accuracy and efficiency remains a challenge in COVID-19 and respiratory disease classification. To develop effective lightweight models, we investigated various optimization strategies, including the convolutional block attention module, depthwise separable convolutions, global average pooling, residual connections, and knowledge distillation. Leveraging these optimization strategies, we introduce two models, LitCovNet 1 and LitCovNet 2, and evaluate them on the COVID-19 Radiography Database and COVID-QU-Ex datasets. LitCovNet 2 achieves 98.78% accuracy on the Radiography Database and 95.64% on COVID-QU-Ex with only 67,349 parameters, while LitCovNet 1 attains 95.70% on COVID-QU-Ex with 131,861 parameters. Despite being 350 \(\times \) smaller than ResNet50, LitCovNet 2 shows comparable accuracy and markedly superior FLOPs, model size, and inference speed. This makes it a perfect candidate for resource-constrained deployment. Benchmarking on multiple edge devices further demonstrates their suitability for resource-constrained deployment. These results establish LitCovNet as an effective lightweight solution for large-scale multiclass COVID-19 and respiratory disease classification.