Scholarlink

<p>Tuberculosis (TB), caused by Mycobacterium tuberculosis, is a major contributor to infectious disease-related mortality. It is the second leading cause of such deaths, following COVID-19. Despite the availability of diagnostic gold standards such as sputum culture and GeneXpert molecular tests, their limited accessibility in low-resource settings poses significant challenges. The World Health Organization (WHO) suggests symptom assessment for TB; however, its accuracy is inadequate. The need for a non-sputum-based, quick, affordable, and extremely accurate diagnostic tool is emphasized in the WHO’s desired product profile for TB triage testing. These requirements are not met by the current TB screening. Given that cough is a primary symptom and transmission method of TB, objective tools to evaluate cough characteristics could enhance TB screening sensitivity. This study aims to explore cough audio frequency as a biomarker for TB detection, leveraging advanced deep-learning techniques to develop a reliable, reproducible cough-based diagnostic model. A lightweight attention-based deep neural network named A-TBNet is proposed for TB detection. Initially, the audio signals are converted into spectrogram images to capture both temporal and spatial characteristics of the audio signal. Further, the spectrogram is processed through the proposed A-TBNet to extract the discriminative features for early detection of TB. A-TBNet model incorporates both channel and pixel attention mechanisms, allowing the model to focus only on the salient features, thus aiding in early convergence of the model. The performance of the A-TBNet is evaluated on the publicly available dataset and achieves the highest accuracy of 86%, proving its effectiveness over state-of-the-art methods. The results highlight the model’s reliability in distinguishing TB-related cough patterns from non-TB cases. The findings suggest that cough-based audio analysis, powered by deep learning, can serve as a potential non-invasive and accessible screening tool for TB. The attention-based architecture enhances the model’s focus on relevant features, aiding in early convergence and accurate TB 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

A-TBNet:attention-guided neural network for tuberculosis detection using cough audio signals

  • 期刊论文

  • 发表日期：

    2026年3月7日

查看全文
  • Geet Sahu,
  • Sumit Kumar,
  • Navya P. Vidhu

摘要

Tuberculosis (TB), caused by Mycobacterium tuberculosis, is a major contributor to infectious disease-related mortality. It is the second leading cause of such deaths, following COVID-19. Despite the availability of diagnostic gold standards such as sputum culture and GeneXpert molecular tests, their limited accessibility in low-resource settings poses significant challenges. The World Health Organization (WHO) suggests symptom assessment for TB; however, its accuracy is inadequate. The need for a non-sputum-based, quick, affordable, and extremely accurate diagnostic tool is emphasized in the WHO’s desired product profile for TB triage testing. These requirements are not met by the current TB screening. Given that cough is a primary symptom and transmission method of TB, objective tools to evaluate cough characteristics could enhance TB screening sensitivity. This study aims to explore cough audio frequency as a biomarker for TB detection, leveraging advanced deep-learning techniques to develop a reliable, reproducible cough-based diagnostic model. A lightweight attention-based deep neural network named A-TBNet is proposed for TB detection. Initially, the audio signals are converted into spectrogram images to capture both temporal and spatial characteristics of the audio signal. Further, the spectrogram is processed through the proposed A-TBNet to extract the discriminative features for early detection of TB. A-TBNet model incorporates both channel and pixel attention mechanisms, allowing the model to focus only on the salient features, thus aiding in early convergence of the model. The performance of the A-TBNet is evaluated on the publicly available dataset and achieves the highest accuracy of 86%, proving its effectiveness over state-of-the-art methods. The results highlight the model’s reliability in distinguishing TB-related cough patterns from non-TB cases. The findings suggest that cough-based audio analysis, powered by deep learning, can serve as a potential non-invasive and accessible screening tool for TB. The attention-based architecture enhances the model’s focus on relevant features, aiding in early convergence and accurate TB detection.