How well does a combined radiomics model distinguish tuberculosis from other lung lesions?

Doctors often use CT scans to look for lung lesions, but it can be hard to tell if a lesion is caused by tuberculosis (TB) or another condition like lung cancer or non-tuberculous mycobacterial lung disease (NTM-LD). A combined radiomics model uses computer analysis of CT images plus patient information to help make this distinction. Research shows these models can be quite accurate, but their performance varies depending on the specific comparison and patient group.

What the research says

A 2025 study developed a combined model that uses both radiomics features (detailed patterns from CT scans) and clinical-semantic features (like patient symptoms and history) to tell TB apart from non-tuberculous solid lung lesions ⁵. This model achieved average precision (AP) scores of 0.91 in the training set and 0.85 in an internal validation set, meaning it correctly identified TB cases most of the time ⁵. However, when tested on a separate group of patients from a later time period (temporal validation), the AP dropped to 0.62, suggesting the model may not perform as well on new patient populations ⁵.

Another study focused on distinguishing TB from NTM-LD, which can look very similar on CT scans ³. A 3D ResNeXt deep learning model achieved an AUC of 0.83 and accuracy of 0.84 on an independent test set of 80 patients ³. This shows that radiomics-based models can help differentiate TB from other mycobacterial infections, though there is still room for improvement.

A separate study created a clinical prediction model (without radiomics) to differentiate NTM-PD from TB using patient characteristics like age, sex, and CT findings such as bronchiectasis and cavitation ⁶. This model had an AUC of 0.84, similar to the radiomics model, suggesting that simpler clinical factors can also be useful ⁶.

For a more complex scenario, a 2025 study used a radiomics model to distinguish patients who have both TB and lung cancer from those with lung cancer alone ⁸. The combined model (radiomics plus clinical features) achieved an AUC of 0.91 in the training set and 0.89 in an internal validation set, showing strong performance ⁸. This highlights that radiomics can help in challenging cases where multiple diseases overlap.

What to ask your doctor

• Would a combined radiomics model be available or helpful for interpreting my CT scan?
• How does the accuracy of a radiomics model compare with other tests like sputum culture or molecular tests for diagnosing TB?
• If a radiomics model suggests TB, what additional tests would you recommend to confirm the diagnosis?
• Are there any limitations to using a radiomics model for my specific type of lung lesion?
• Could a clinical prediction model (using my age, symptoms, and CT features) be as useful as a radiomics model in my case?