Scholarlink

Tuberculosis (TB) continues to be a significant threat to global health, especially within low-resource environments where the availability of skilled radiologists is limited. In response, we introduce TUBER-ENSEMBLE, a new deep learning framework that operates on an ensemble basis. This framework integrates three distinct CNN architectures—ResNet50, DenseNet121, and InceptionV3—to automate the detection of TB from chest X-ray images. Tested on the Kaggle TB dataset of 4200 images, our ensemble model, which uses a soft voting mechanism, reached an accuracy of 98.45%, precision of 98%, and recall of 92%, notably surpassing the performance of the individual models. Among the single models, InceptionV3 yielded the strongest results with 98% accuracy and a 97% F1-score. Conversely, ResNet50 displayed lower sensitivity, achieving 83.6% accuracy and a 47% F1-score. The proposed ensemble method exhibits enhanced generalization and robustness, presenting a viable and scalable tool for TB screening initiatives in regions that lack sufficient diagnostic expertise.

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

TUBER-ENSEMBLE: An Ensemble of Deep Convolutional Neural Networks for Tuberculosis Detection from Chest X-Rays

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mohssine Kissane,
  • Amine Zeguendry,
  • Walid Ahkouk,
  • Abdelilah Dahou

摘要

Tuberculosis (TB) continues to be a significant threat to global health, especially within low-resource environments where the availability of skilled radiologists is limited. In response, we introduce TUBER-ENSEMBLE, a new deep learning framework that operates on an ensemble basis. This framework integrates three distinct CNN architectures—ResNet50, DenseNet121, and InceptionV3—to automate the detection of TB from chest X-ray images. Tested on the Kaggle TB dataset of 4200 images, our ensemble model, which uses a soft voting mechanism, reached an accuracy of 98.45%, precision of 98%, and recall of 92%, notably surpassing the performance of the individual models. Among the single models, InceptionV3 yielded the strongest results with 98% accuracy and a 97% F1-score. Conversely, ResNet50 displayed lower sensitivity, achieving 83.6% accuracy and a 47% F1-score. The proposed ensemble method exhibits enhanced generalization and robustness, presenting a viable and scalable tool for TB screening initiatives in regions that lack sufficient diagnostic expertise.