Scholarlink

Purpose <p>Pulmonary tuberculosis remains a major public health issue in Indonesia, intensified by a shortage of radiologists for early detection. Although deep learning enables automated screening, most current models focus on binary classification and often miss post-TB sequelae, a critical factor in elimination strategies. This study aimed to evaluate the clinical implementation of a deep learning model based on the Xception architecture for multiclass tuberculosis detection in chest X-ray (CXR) images, including normal, active, and post-TB sequelae tuberculosis.</p> Methods <p>This analytical observational study evaluated a prospective implementation cohort of 529 adult chest X-ray images from five regions in Central Java, Indonesia, comprising 259 normal, 246 active TB, and 24 post-TB sequelae cases. The Xception model was developed using a separate dataset, and its clinical performance was subsequently evaluated in an independent implementation cohort. Agreement with radiologist interpretations was assessed using Cohen’s kappa, and user acceptance was evaluated by 30 radiologists using the User Experience Questionnaire Plus (UEQ+).</p> Results <p>In the prospective implementation cohort, the model showed high multiclass classification performance, with an overall accuracy of 98.95%, a macro-average precision of 96.30%, and a macro-average recall of 96.29%. The model demonstrated high sensitivity for active tuberculosis (recall 97.97%) and high precision for post-TB sequelae (precision 91.67%), although recall for post-TB sequelae cases remained lower due to overlapping radiographic features. Agreement with radiologist interpretations was high, and user evaluation indicated very good perceived system quality, although hardware security remained the main area for improvement.</p> Conclusions <p>The implemented deep learning system showed high agreement with radiologist interpretations and favorable user acceptance, suggesting its potential as a screening support tool for classifying normal, active TB, and post-TB sequelae on CXR. External validation across a diverse population is required in future work to establish generalizability.</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

Clinical implementation of an Xception-based deep learning system for multiclass tuberculosis detection on chest X-ray images

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月22日

查看全文
  • Darmini Darmini,
  • Bambang Budi Raharjo,
  • Mahalul Azam,
  • Intan Zainafree,
  • Anan Nugroho,
  • Rafik Kladius,
  • Irfan Alfian Rizqi

摘要

Purpose

Pulmonary tuberculosis remains a major public health issue in Indonesia, intensified by a shortage of radiologists for early detection. Although deep learning enables automated screening, most current models focus on binary classification and often miss post-TB sequelae, a critical factor in elimination strategies. This study aimed to evaluate the clinical implementation of a deep learning model based on the Xception architecture for multiclass tuberculosis detection in chest X-ray (CXR) images, including normal, active, and post-TB sequelae tuberculosis.

Methods

This analytical observational study evaluated a prospective implementation cohort of 529 adult chest X-ray images from five regions in Central Java, Indonesia, comprising 259 normal, 246 active TB, and 24 post-TB sequelae cases. The Xception model was developed using a separate dataset, and its clinical performance was subsequently evaluated in an independent implementation cohort. Agreement with radiologist interpretations was assessed using Cohen’s kappa, and user acceptance was evaluated by 30 radiologists using the User Experience Questionnaire Plus (UEQ+).

Results

In the prospective implementation cohort, the model showed high multiclass classification performance, with an overall accuracy of 98.95%, a macro-average precision of 96.30%, and a macro-average recall of 96.29%. The model demonstrated high sensitivity for active tuberculosis (recall 97.97%) and high precision for post-TB sequelae (precision 91.67%), although recall for post-TB sequelae cases remained lower due to overlapping radiographic features. Agreement with radiologist interpretations was high, and user evaluation indicated very good perceived system quality, although hardware security remained the main area for improvement.

Conclusions

The implemented deep learning system showed high agreement with radiologist interpretations and favorable user acceptance, suggesting its potential as a screening support tool for classifying normal, active TB, and post-TB sequelae on CXR. External validation across a diverse population is required in future work to establish generalizability.