Scholarlink

<p>COVID-19, an acute and extremely infectious respiratory disease, has taken the lives of millions worldwide. For early identification, Computed Tomography (CT) scans have become popular over RT-PCR tests. Numerous deep-learning approaches have been presented by extensive research in this discipline. However, the majority of present models comprise a vast amount of parameters and increased complexity, making it challenging to train and operate them on resource-constrained devices. This research aims to reduce the computational complexity while maintaining reliability and performance. This article proposes a lightweight Convolutional Neural Network (CNN) aided by an attention mechanism that is capable of reliably identifying COVID-19 CT images. Having significantly fewer parameters, it also addresses the challenge of integration on devices with limited resources, thus improving the early diagnosis of COVID-19. The experiments utilized two well-recognized datasets, the COVIDx CT benchmark dataset and the Large COVID-19 CT-scan slice dataset. To ensure unbiased results, five-fold cross-validation, ROC curve, etc., were employed. The proposed architecture, with a parameter count of less than 5 million, achieved accuracies of 98.50% and 98.13% for the two datasets, respectively. Remarkably, it only took 0.25&#xa0;s to classify a preprocessed image on a medium-end device without GPU support. In addition, Grad-CAM and SHAP visualizations validate the explainability of the proposed approach. The experimental results demonstrate superior generalization as well as efficacy of the proposed method, despite having a little computational overhead. Hence, if implemented appropriately, this study has the potential to significantly aid in CT-based COVID-19 detection.</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

Lightweight COVID-19 Detection from Chest CT-Scans Using Attention-Based CNN

  • 期刊论文

  • 发表日期：

    2025年9月25日

查看全文
  • Ananta Raha

摘要

COVID-19, an acute and extremely infectious respiratory disease, has taken the lives of millions worldwide. For early identification, Computed Tomography (CT) scans have become popular over RT-PCR tests. Numerous deep-learning approaches have been presented by extensive research in this discipline. However, the majority of present models comprise a vast amount of parameters and increased complexity, making it challenging to train and operate them on resource-constrained devices. This research aims to reduce the computational complexity while maintaining reliability and performance. This article proposes a lightweight Convolutional Neural Network (CNN) aided by an attention mechanism that is capable of reliably identifying COVID-19 CT images. Having significantly fewer parameters, it also addresses the challenge of integration on devices with limited resources, thus improving the early diagnosis of COVID-19. The experiments utilized two well-recognized datasets, the COVIDx CT benchmark dataset and the Large COVID-19 CT-scan slice dataset. To ensure unbiased results, five-fold cross-validation, ROC curve, etc., were employed. The proposed architecture, with a parameter count of less than 5 million, achieved accuracies of 98.50% and 98.13% for the two datasets, respectively. Remarkably, it only took 0.25 s to classify a preprocessed image on a medium-end device without GPU support. In addition, Grad-CAM and SHAP visualizations validate the explainability of the proposed approach. The experimental results demonstrate superior generalization as well as efficacy of the proposed method, despite having a little computational overhead. Hence, if implemented appropriately, this study has the potential to significantly aid in CT-based COVID-19 detection.