Scholarlink

The correct interpretation of chest radiographs (CXR) still remains a challenge in the clinical practice, particularly in a mass emergency such as the COVID-19 pandemic, where a fast and accurate approach to diagnosis is of utmost importance. In the proposed work, an on-segmentation pipeline based on integrated classification combines four deep CNN models. (DenseNet121, ResNet50, EfficientNetB0, and ConvNeXt-Tiny) on the classification of various. Images on CXR to four categories COVID-19, lung opacities, viral pneumonia and normal. We introduce a novel hybrid architecture that integrates DenseNet121’s dense connectivity with U-Net’s encoder–decoder framework, enhanced with channel-wise attention mechanisms for improved spatial feature learning. Developed a balanced data set of each class had 5380 images and 1345 samples to represent fair model learning. All models were trained using a steady stream of data–data augmentation using Albumentations, early stopping and a maximum of 50 epochs. Our comprehensive evaluation includes accuracy, precision, recall, F1-score, IoU, Dice coefficient, statistical significance testing (Wilcoxon signed—rank test, p < 0.05), and computational efficiency analysis (training time, inference speed, parameter count). Its high connectivity and the effective re-use of features of the DenseNet121 model made it more performance better thus it was not hard to apply it to other datasets. Our hybrid DenseNet121 + U-Net model achieved superior performance with 98.05% validation accuracy (95% CI: 96.8–99.3%, p < 0.001 versus baseline CNN), mean IoU of 0.847 ± 0.032, and Dice coefficient of 0.891 ± 0.028. For improved interpretability, we implement a multi-modal explainability framework combining Grad-CAM, Score-CAM, and SHAP analysis with heatmaps focusing on lung regions affecting prediction outcomes that correspond to radiologic signs. In summary, this work presents a robust and interpretable deep learning solution that integrates classification and segmentation to enable enhanced clinical decision support in thoracic imaging.

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

Explainable Deep Learning for COVID-19 and Chest Disease Detection: A Dual-Model Approach Using DenseNet121 and UNet

  • 章节

  • 发表日期：

    2026年

查看全文
  • T. Grace Shalini,
  • S. S. Krishikaa Mathi Bharathi,
  • T. Padmapriya

摘要

The correct interpretation of chest radiographs (CXR) still remains a challenge in the clinical practice, particularly in a mass emergency such as the COVID-19 pandemic, where a fast and accurate approach to diagnosis is of utmost importance. In the proposed work, an on-segmentation pipeline based on integrated classification combines four deep CNN models. (DenseNet121, ResNet50, EfficientNetB0, and ConvNeXt-Tiny) on the classification of various. Images on CXR to four categories COVID-19, lung opacities, viral pneumonia and normal. We introduce a novel hybrid architecture that integrates DenseNet121’s dense connectivity with U-Net’s encoder–decoder framework, enhanced with channel-wise attention mechanisms for improved spatial feature learning. Developed a balanced data set of each class had 5380 images and 1345 samples to represent fair model learning. All models were trained using a steady stream of data–data augmentation using Albumentations, early stopping and a maximum of 50 epochs. Our comprehensive evaluation includes accuracy, precision, recall, F1-score, IoU, Dice coefficient, statistical significance testing (Wilcoxon signed—rank test, p < 0.05), and computational efficiency analysis (training time, inference speed, parameter count). Its high connectivity and the effective re-use of features of the DenseNet121 model made it more performance better thus it was not hard to apply it to other datasets. Our hybrid DenseNet121 + U-Net model achieved superior performance with 98.05% validation accuracy (95% CI: 96.8–99.3%, p < 0.001 versus baseline CNN), mean IoU of 0.847 ± 0.032, and Dice coefficient of 0.891 ± 0.028. For improved interpretability, we implement a multi-modal explainability framework combining Grad-CAM, Score-CAM, and SHAP analysis with heatmaps focusing on lung regions affecting prediction outcomes that correspond to radiologic signs. In summary, this work presents a robust and interpretable deep learning solution that integrates classification and segmentation to enable enhanced clinical decision support in thoracic imaging.