Scholarlink

This paper evaluates the dual hybrid convolutional neural networks of Xception-DenseNet121 and Xception-ResNet50 for the classification of malaria across blood smear images. Both networks were early-stopped and experimented for performance over ten epochs. Training accuracy of Xception-DenseNet121 achieved 93.96%, but validation loss had sudden lows and peaks, which indicate overfitting. In comparison, Xception-ResNet50 had a stable validation accuracy of 95.74%, quick convergence, and a negligible final loss of 0.1282 to indicate better generalization. Confusion matrices had fewer errors and higher-class balance for Xception-ResNet50. ROC curves and AUC measures (0.9905 in comparison to 0.9820) validated the observation. Visualization of the saliency map confirmed that both models had the appropriate attention in infection regions, enhancing interpretability. Further, Xception-ResNet50 achieved higher F1-scores (0.96 for Parasitized, 0.96 for Uninfected) compared to the performance of Xception-DenseNet121 (0.94 for Parasitized, 0.94 for Uninfected), validating classification consistency. Cross-validation also had less variance in performance measures for Xception-ResNet50. The results validate the better performance of the automation-based diagnosis system for malaria in the form of Xception-ResNet50.

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 Hybrid Convolutional Networks for Accurate Malaria Diagnosis in Blood Smear Images

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Mohammad Mehedi Hasan Munna,
  • Nakib Aman Turzo,
  • Smita Saha,
  • Toriqul Islam,
  • Iqbal Hossain Safy,
  • Fairuj Saima,
  • Sabiha Nusrat

摘要

This paper evaluates the dual hybrid convolutional neural networks of Xception-DenseNet121 and Xception-ResNet50 for the classification of malaria across blood smear images. Both networks were early-stopped and experimented for performance over ten epochs. Training accuracy of Xception-DenseNet121 achieved 93.96%, but validation loss had sudden lows and peaks, which indicate overfitting. In comparison, Xception-ResNet50 had a stable validation accuracy of 95.74%, quick convergence, and a negligible final loss of 0.1282 to indicate better generalization. Confusion matrices had fewer errors and higher-class balance for Xception-ResNet50. ROC curves and AUC measures (0.9905 in comparison to 0.9820) validated the observation. Visualization of the saliency map confirmed that both models had the appropriate attention in infection regions, enhancing interpretability. Further, Xception-ResNet50 achieved higher F1-scores (0.96 for Parasitized, 0.96 for Uninfected) compared to the performance of Xception-DenseNet121 (0.94 for Parasitized, 0.94 for Uninfected), validating classification consistency. Cross-validation also had less variance in performance measures for Xception-ResNet50. The results validate the better performance of the automation-based diagnosis system for malaria in the form of Xception-ResNet50.