Diagnostic Performance of Artificial Intelligence–Based Methods for Tuberculosis Detection: Systematic Review

Seng Hansun, Ahmadreza Argha, Ivan Bakhshayeshi, Arya Wicaksana, Hamid Alinejad‐Rokny, Greg J. Fox, Siaw‐Teng Liaw, Branko G. Celler, Guy B. MarksJournal of Medical Internet Research 2025 · DOI

TL;DR

This systematic review synthesizes the diagnostic performance of AI methods for tuberculosis (TB) detection, finding that AI, particularly deep learning (DL) on radiographic data, shows high accuracy. However, the field needs more research on domain-shift analysis to ensure real-world applicability.

What Problem Does This Paper Address?

Tuberculosis (TB) remains a major global health concern, and current diagnostic tools have limitations. While Artificial Intelligence (AI) offers promising solutions, there's a need for a comprehensive understanding of its diagnostic performance across various AI techniques, data modalities, and reference standards. This paper addresses this by systematically reviewing and evaluating the existing literature on AI-based TB detection.

Methods

The authors conducted a systematic review following PRISMA 2020 guidelines, searching three major databases (Scopus, PubMed, ACM Digital Library). They included 152 studies and assessed their risk of bias using the QUADAS-2 tool. The review synthesized findings on AI methods, data types, evaluation techniques, and performance metrics. A narrative synthesis approach was used due to the heterogeneity of the studies, making a meta-analysis impractical.

Datasets

The paper analyzed studies that utilized various datasets. Prominent among these were the Shenzhen (SZ) and Montgomery County (MC) datasets, both publicly available chest X-ray (CXR) datasets from the US National Library of Medicine. Other datasets mentioned include Kaggle TB CXR and TBX11K. The paper notes that while these datasets are widely used, their limited size can pose challenges for DL models.

Key Insights and Ideas

  • DL Dominance: Deep learning, especially Convolutional Neural Networks (CNNs), is the most prevalent AI approach for TB detection, outperforming traditional Machine Learning (ML) in many aspects.
  • Radiographic Data is King: Radiographic data, particularly chest X-rays (CXRs), is the most common data modality used, likely due to its availability and the success of DL models in image analysis.
  • Single Modality Preference: The majority of studies used a single data modality, which, surprisingly, tended to show higher median performance metrics and narrower distributions compared to multimodal approaches in this review. However, the authors caution that this might not hold true under identical experimental conditions.
  • Transfer Learning's Advantage: Transfer learning (TL) is a widely adopted strategy, particularly with DL, and consistently shows a higher median accuracy and narrower distribution range, indicating its effectiveness in handling limited data volumes.
  • The Domain-Shift Gap: A significant limitation identified is the lack of research on domain-shift analysis. Most studies evaluate AI models on data similar to their training data (in-domain), which may not reflect real-world scenarios where data distributions can differ significantly, potentially leading to performance degradation.

Key Results

  • High Overall Performance: AI methods demonstrated strong diagnostic performance across the included studies, with mean accuracy of 91.93%, mean AUC of 93.48%, mean sensitivity of 92.77%, and mean specificity of 92.39%.
  • DL Architectures: Popular DL architectures included VGG-16, ResNet-50, and DenseNet-121.
  • ML Techniques: Among ML methods, Support Vector Machines (SVMs), Random Forests (RFs), and Logistic Regression (LR) were most common.
  • Biomarker Performance: Radiographic biomarkers generally showed superior performance compared to molecular/biochemical and physiological biomarkers.
  • Reference Standard Impact: Studies using human reader reference standards tended to perform better than those using bacteriological or combined standards.
  • Risk of Bias: A notable portion of studies (11.8%) were flagged for a high risk of bias, primarily related to patient selection, index test, and reference standards, highlighting areas for improvement in study design.

Limitations and Future Work

The authors acknowledge several limitations: the search was restricted to English-language publications from three databases, potentially missing relevant studies. They also assumed uniformity across radiographic datasets, which might overlook variations. The decision not to perform a meta-analysis due to study heterogeneity is noted.

The authors strongly recommend future research to prioritize domain-shift analyses. This is crucial for understanding how AI models perform when applied to data from different sources or settings than those they were trained on, which is essential for real-world deployment. Further studies exploring multimodal approaches under similar experimental conditions and investigating emerging DL architectures are also encouraged.

Summary updated

Abstract

PROSPERO CRD42023453611; https://www.crd.york.ac.uk/PROSPERO/view/CRD42023453611.
BibTeX
@article{hansun2025diagnostic,
  author = {Seng Hansun and Ahmadreza Argha and Ivan Bakhshayeshi and Arya Wicaksana and Hamid Alinejad‐Rokny and Greg J. Fox and Siaw‐Teng Liaw and Branko G. Celler and Guy B. Marks},
  title = {{Diagnostic Performance of Artificial Intelligence–Based Methods for Tuberculosis Detection: Systematic Review}},
  journal = {Journal of Medical Internet Research},
  year = {2025},
  doi = {10.2196/69068},
}
Figure