Scholarlink

<p>COVID-19 has affected over 600 million people globally, leading to a surge in demand for rapid and accurate diagnostic tools, especially for respiratory conditions like COVID-19, pneumonia, and lung opacity. Chest X-ray (CXR) imaging, being widely available and cost-effective, has been extensively used for diagnosing such diseases. However, due to poor feature extraction and localization techniques, conventional diagnostic methods need help with multi-class disease classification from CXR images. This research addresses the gap by introducing an advanced model that improves accuracy and efficiency in identifying multiple respiratory conditions from CXR images. So, this work focuses on developing a deep CXR image analysis network (DICA-Net) to classify four significant classes: COVID-19, lung opacity, pneumonia viral, and normal. The DICA-Net methodology comprises four essential procedures, beginning with data preprocessing to enhance image quality and consistency. Subsequently, DICA-Net introduces an Attention U-Net (AU-Net) segmentation approach to precisely delineate disease regions, allowing for more accurate disease localization. Then, pigeon local search optimization (PLSO) was developed for feature extraction and selection, a powerful technique for identifying relevant diagnostic features within the CXR images. The PLSO enhances the interpretability of the model while preserving its diagnostic accuracy. Finally, a custom convolutional neural network (CCNN) classifier is employed to classify the four disease classes, harnessing the rich feature representations extracted through the previous steps. The DICA-Net model demonstrates exceptional performance on multi-class CXR dataset, achieving overall accuracy of 99.211%, precision of 99.087%, recall of 99.407%, specificity of 99.23%, and F-measure of 99.184%. This integrated approach can significantly improve the efficiency, accuracy, and accessibility of CXR-based disease identification, addressing challenges related to expert shortages and diagnostic delays.</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

DICA-Net: optimizing chest X-ray classification with attention U-Net and pigeon local search

  • 期刊论文

  • 发表日期：

    2025年4月9日

查看全文
  • G. S. V. R. Abhishek,
  • Amit Singla,
  • D. Eshwar

摘要

COVID-19 has affected over 600 million people globally, leading to a surge in demand for rapid and accurate diagnostic tools, especially for respiratory conditions like COVID-19, pneumonia, and lung opacity. Chest X-ray (CXR) imaging, being widely available and cost-effective, has been extensively used for diagnosing such diseases. However, due to poor feature extraction and localization techniques, conventional diagnostic methods need help with multi-class disease classification from CXR images. This research addresses the gap by introducing an advanced model that improves accuracy and efficiency in identifying multiple respiratory conditions from CXR images. So, this work focuses on developing a deep CXR image analysis network (DICA-Net) to classify four significant classes: COVID-19, lung opacity, pneumonia viral, and normal. The DICA-Net methodology comprises four essential procedures, beginning with data preprocessing to enhance image quality and consistency. Subsequently, DICA-Net introduces an Attention U-Net (AU-Net) segmentation approach to precisely delineate disease regions, allowing for more accurate disease localization. Then, pigeon local search optimization (PLSO) was developed for feature extraction and selection, a powerful technique for identifying relevant diagnostic features within the CXR images. The PLSO enhances the interpretability of the model while preserving its diagnostic accuracy. Finally, a custom convolutional neural network (CCNN) classifier is employed to classify the four disease classes, harnessing the rich feature representations extracted through the previous steps. The DICA-Net model demonstrates exceptional performance on multi-class CXR dataset, achieving overall accuracy of 99.211%, precision of 99.087%, recall of 99.407%, specificity of 99.23%, and F-measure of 99.184%. This integrated approach can significantly improve the efficiency, accuracy, and accessibility of CXR-based disease identification, addressing challenges related to expert shortages and diagnostic delays.