Scholarlink

<p>Coronavirus Disease 2019 (COVID-19) has rapidly evolved into a global public health crisis. While Reverse Transcription Polymerase Chain Reaction (RT-PCR) remains the widely accepted standard for COVID-19 diagnosis, its relatively slow turnaround time can delay critical clinical decisions, potentially worsening patient outcomes, including the development of COVID-19-induced pneumonia. In response, deep learning techniques have shown strong potential for the rapid analysis of lung ultrasound (LUS) images, offering a non-invasive and faster alternative for assessment. This study presents an integration of a Hierarchical Spatial Transformer Network (HSTN) with a transformed VGG-16 model for predicting COVID-19 severity in LUS videos. The HSTN module is employed to localize pathological artifacts associated with COVID-19 in LUS frames, improving feature representation. Subsequently, a fine-tuned VGGNet-16 model assigns severity scores to individual frames. These frame-level scores are then aggregated using a uninorm-based function to compute a video-level severity score, categorizing patients as healthy, COVID-19 positive, or pneumonia cases. An experimental evaluation on the ICLUS-DB dataset demonstrates that the integrated framework achieves a severity classification accuracy of 93.27%, outperforming several existing approaches in terms of diagnostic performance, while highlighting the value of combining spatial localization and video-level aggregation for LUS-based COVID-19 assessment.</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

Confluence of HSTN and Transformed VGGNet-16 Model: A Deep Learning Assisted Robust COVID-19 Severity Detection Approach in Lung Ultrasound

  • 期刊论文

  • 发表日期：

    2025年12月9日

查看全文
  • B. Selvakumar,
  • P. Kaviya

摘要

Coronavirus Disease 2019 (COVID-19) has rapidly evolved into a global public health crisis. While Reverse Transcription Polymerase Chain Reaction (RT-PCR) remains the widely accepted standard for COVID-19 diagnosis, its relatively slow turnaround time can delay critical clinical decisions, potentially worsening patient outcomes, including the development of COVID-19-induced pneumonia. In response, deep learning techniques have shown strong potential for the rapid analysis of lung ultrasound (LUS) images, offering a non-invasive and faster alternative for assessment. This study presents an integration of a Hierarchical Spatial Transformer Network (HSTN) with a transformed VGG-16 model for predicting COVID-19 severity in LUS videos. The HSTN module is employed to localize pathological artifacts associated with COVID-19 in LUS frames, improving feature representation. Subsequently, a fine-tuned VGGNet-16 model assigns severity scores to individual frames. These frame-level scores are then aggregated using a uninorm-based function to compute a video-level severity score, categorizing patients as healthy, COVID-19 positive, or pneumonia cases. An experimental evaluation on the ICLUS-DB dataset demonstrates that the integrated framework achieves a severity classification accuracy of 93.27%, outperforming several existing approaches in terms of diagnostic performance, while highlighting the value of combining spatial localization and video-level aggregation for LUS-based COVID-19 assessment.