TB Research
← Back to search

Diagnostic accuracy of CT-based radiomics models in differentiating lung cancer from tuberculosis in pulmonary lesions: a systematic review and meta-analysis.

Hadi Sahrai, Jamal Behnood, Mansoureh Baradaran, Amirreza Khalaji, Ali Norouzi, Farzaneh Shojaeshafiei, Seyedeh Mahdieh Seyed Ebrahimi, Sanam Mohammadzadeh, et al. (10 authors)

BMC cancer · 2025-12

Abstract

BACKGROUND: Distinguishing lung cancer (LC) from pulmonary tuberculosis (TB) on CT is difficult. We synthesized evidence on radiomics, clinical, and combined models for LC–TB discrimination.

METHODS: PubMed, Web of Science, Embase, and Scopus were searched to August 2025 following PRISMA. Metrics were harmonized to LC-positive/TB-negative and pooled with a bivariate random-effects model. Study quality was assessed with QUADAS-2 and METRICS.

RESULTS: Fourteen retrospective studies (4281 participants) were included. Radiomics models (11 validations cohorts) achieved pooled sensitivity 0.80 (95% CI 0.74&#x2013;0.86) and specificity 0.83 (0.75&#x2013;0.88); SROC AUC 0.88 (0.85&#x2013;0.91). At a 25% pre-test probability, radiomics models corresponded to post-test probabilities of 61% and 7%. Deeks&#x2019; funnel asymmetry test showed no small-study effects (&#x2009;=&#x2009;0.88). Clinical-only models performed more modestly (sensitivity 0.60, specificity 0.80, AUC 0.77). Combined radiomics&#x2009;+&#x2009;clinical models performed best (sensitivity 0.82, specificity 0.93, AUC 0.90). Head-to-head comparison showed higher specificity for radiomics versus clinical models (&#x2009;=&#x2009;0.02), and higher sensitivity for combined models versus radiomics (&#x2009;<&#x2009;0.001) without a clear specificity difference (&#x2009;=&#x2009;0.41). Prespecified subgroup analyses indicated that models retaining&#x2009;>&#x2009;10 radiomic features, developed in cohorts restricted to nodules&#x2009;<&#x2009;3&#xa0;cm, and using non-contrast CT tended to perform better, whereas externally validated cohorts showed lower accuracy than internal test sets, and both nodule-size spectrum and CT acquisition phase emerged as major contributors to between-study heterogeneity.

CONCLUSIONS: CT-based radiomics adds discriminative value beyond clinical variables, and integrating clinical information with radiomics yields the most favorable overall accuracy while highlighting the importance of broader external validation.

SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s12885-025-15446-5.