Scholarlink

<p>As COVID-19 is a pneumonia disease it mostly affects the lungs of patients, also the outbreak of COVID-19 disease is about to drop but not completely. So the main concern for researchers and clinicians is to identify COVID-19 from other pneumonia diseases as usual processes such as from chest X-Rays instead of antigen or PCR tests. However medical imaging on X-Rays has shown promising results using machine learning, and deep learning-based techniques. A deep learning-based model requires a large amount of data which is not possible for chest X-Ray of COVID-19 patients, so transfer learning can be employed in this case. Moreover, the problem of transfer learning is the uncertainty in model structure which leads to biased results instead of promising results. This study intends to reduce the uncertainty in transfer learning models used to identify COVID-19, or other pneumonia diseases from chest X-rays, with an emphasis on optimizing model structure for improved accuracy. A novel genetic optimization method has been proposed in our work to improve transfer learning by optimizing model architectures and freezing layers inside the VGG16 architecture, because of its ease of understanding in structural architectures. We use a relatively small dataset to empirically compare the method to typical transfer learning and other optimization approaches. Our genetic transfer learning method obtained 97.67% accuracy in identifying COVID-19 using VGG16. This technique outperformed the straightforward counterpart of its transfer learning approach by an average accuracy of 0.98%. In comparison to other models that use the same dataset as our work, our best model outperforms them by an average precision of 7.13%, an average recall of 7.38%, an average f1-score of 7.41%, and an average accuracy of 7.73%. Similarly, our best model outperforms other deep learning-based works that use small datasets by an average accuracy of 4.47%. Also, our work implies that there is an additional opportunity for improvement in classification accuracy through further exploitation of the best model found using our genetic optimization technique.</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

Genetic Optimization-Based Layers Tuning and Freezing in Deep CNN for Low-Cost Disease Detection Using Chest X-Rays

  • 期刊论文

  • 发表日期：

    2025年6月16日

查看全文
  • Md. Jahidul Islam

摘要

As COVID-19 is a pneumonia disease it mostly affects the lungs of patients, also the outbreak of COVID-19 disease is about to drop but not completely. So the main concern for researchers and clinicians is to identify COVID-19 from other pneumonia diseases as usual processes such as from chest X-Rays instead of antigen or PCR tests. However medical imaging on X-Rays has shown promising results using machine learning, and deep learning-based techniques. A deep learning-based model requires a large amount of data which is not possible for chest X-Ray of COVID-19 patients, so transfer learning can be employed in this case. Moreover, the problem of transfer learning is the uncertainty in model structure which leads to biased results instead of promising results. This study intends to reduce the uncertainty in transfer learning models used to identify COVID-19, or other pneumonia diseases from chest X-rays, with an emphasis on optimizing model structure for improved accuracy. A novel genetic optimization method has been proposed in our work to improve transfer learning by optimizing model architectures and freezing layers inside the VGG16 architecture, because of its ease of understanding in structural architectures. We use a relatively small dataset to empirically compare the method to typical transfer learning and other optimization approaches. Our genetic transfer learning method obtained 97.67% accuracy in identifying COVID-19 using VGG16. This technique outperformed the straightforward counterpart of its transfer learning approach by an average accuracy of 0.98%. In comparison to other models that use the same dataset as our work, our best model outperforms them by an average precision of 7.13%, an average recall of 7.38%, an average f1-score of 7.41%, and an average accuracy of 7.73%. Similarly, our best model outperforms other deep learning-based works that use small datasets by an average accuracy of 4.47%. Also, our work implies that there is an additional opportunity for improvement in classification accuracy through further exploitation of the best model found using our genetic optimization technique.