Scholarlink

<p>The COVID-19 pandemic has had a devastating impact on the world. Therefore, early and accurate detection and diagnosis are essential to prevent the spread of the virus. We hypothesize that the integrated approach of the feature fusion-based ensemble (FFBE) and the weighted average-based ensemble (WABE) is superior to the FFBE and the WABE models derived using transfer learning (TL) models, namely VGG16, VGG19, DenseNet169, and DenseNet121. Using the hybrid deep learning architecture VGG-UNet for data segmentation, our proposed cloud-based model COVLIAS 3.0<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(_{\text {Fusion}}\)</EquationSource> </InlineEquation> (Global Biomedical Technologies, Inc., Roseville, CA, USA) has been assessed on five types of data combination (DC1, DC2, DC3, DC4 and DC5) that uses CroMed (COVID-19), NovMED (COVID-19), and NovMED (Control) data sets. COVLIAS 3.0<InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(_{\text {Fusion}}\)</EquationSource> </InlineEquation> is the integration of the FFBE and WABE methods that captures an exhaustive data representation, while the WABE aggregates predictions from multiple FFBE models, thus improving accuracy and dependability. COVLIAS 3.0<InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(_{\text {Fusion}}\)</EquationSource> </InlineEquation> has also been utilized to perform unseen data analysis, leading to generalization. Lastly, we show the heatmaps as part of the explainability of the FFBE models. The COVLIAS 3.0<InlineEquation ID="IEq6"> <EquationSource Format="TEX">\(_{\text {Fusion}}\)</EquationSource> </InlineEquation> integrated ensemble model has attained a remarkable 4.6% better accuracy than the WABE method and 4.8% better accuracy than the FFBE, and outperformed the latest published model’s mean accuracy by 4.17% using DC1. Our best results using DC1 have shown a mean accuracy, AUC, precision, F1 score, and recall of 99.35%, 0.99 (p&lt;0.0001), 100%, 100%, and 99%, respectively. Using the complementary strengths of FFBE and WABE, our model achieves better accuracy, robustness, and generalization compared to those of FFBE and WABE alone, respectively.</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

Covlias 3.0Fusion: a novel deep learning ensemble strategy for COVID-19 diagnosis in computed tomography scans

  • 期刊论文

  • 发表日期：

    2025年11月21日

查看全文
  • Arun Kumar Dubey,
  • Achin Jain,
  • Meenakshi Gupta,
  • Neera Aggarwal,
  • Luca Saba,
  • Jasjit S. Suri

摘要

The COVID-19 pandemic has had a devastating impact on the world. Therefore, early and accurate detection and diagnosis are essential to prevent the spread of the virus. We hypothesize that the integrated approach of the feature fusion-based ensemble (FFBE) and the weighted average-based ensemble (WABE) is superior to the FFBE and the WABE models derived using transfer learning (TL) models, namely VGG16, VGG19, DenseNet169, and DenseNet121. Using the hybrid deep learning architecture VGG-UNet for data segmentation, our proposed cloud-based model COVLIAS 3.0 \(_{\text {Fusion}}\) (Global Biomedical Technologies, Inc., Roseville, CA, USA) has been assessed on five types of data combination (DC1, DC2, DC3, DC4 and DC5) that uses CroMed (COVID-19), NovMED (COVID-19), and NovMED (Control) data sets. COVLIAS 3.0 \(_{\text {Fusion}}\) is the integration of the FFBE and WABE methods that captures an exhaustive data representation, while the WABE aggregates predictions from multiple FFBE models, thus improving accuracy and dependability. COVLIAS 3.0 \(_{\text {Fusion}}\) has also been utilized to perform unseen data analysis, leading to generalization. Lastly, we show the heatmaps as part of the explainability of the FFBE models. The COVLIAS 3.0 \(_{\text {Fusion}}\) integrated ensemble model has attained a remarkable 4.6% better accuracy than the WABE method and 4.8% better accuracy than the FFBE, and outperformed the latest published model’s mean accuracy by 4.17% using DC1. Our best results using DC1 have shown a mean accuracy, AUC, precision, F1 score, and recall of 99.35%, 0.99 (p<0.0001), 100%, 100%, and 99%, respectively. Using the complementary strengths of FFBE and WABE, our model achieves better accuracy, robustness, and generalization compared to those of FFBE and WABE alone, respectively.