Combined radiomics-clinical model improves AP for differentiating pulmonary tuberculosis from non-tuberculous lung lesions on CT

This observational study assessed the feasibility of a combined model integrating radiomics and clinical-semantic features for differentiating pulmonary tuberculosis from non-tuberculous solid lung lesions using contrast-enhanced CT. The analysis included 900 patients enrolled prior to October 2017. The primary outcome measured average precision (AP) across three datasets: training, internal validation, and temporal validation.

In the training set, the combined model achieved an AP of 0.91, compared to 0.64 for the clinical-semantic model alone. In the internal validation set, the combined model maintained an AP of 0.85 versus 0.61 for the comparator. However, performance dropped significantly in the temporal validation set, where the combined model yielded an AP of 0.62 compared to 0.41 for the clinical-semantic model.

Safety data, including adverse events, discontinuations, and tolerability, were not reported. The study limitations are not explicitly detailed in the provided text, though the substantial drop in AP from internal to temporal validation sets implies potential overfitting or lack of generalizability over time. Funding sources and conflicts of interest were not reported.

The practice relevance remains uncertain given the observational design and the decline in model performance over time. Clinicians should interpret these results with caution, noting that the combined approach may not yet be ready for routine deployment without further external validation.