Daniel Capellán Martín, first author and researcher at UPM, explains: "We designed this tool to be extremely efficient without sacrificing precision or performance, with the goal of making it suitable even for mobile devices. This could help bring tuberculosis diagnosis closer to rural areas with high disease incidence, where resources and access to specialized radiologists are very limited”.

 

Juan José Gómez Valverde, second author of the study and professor at UPM, adds: “Pretraining on adult data allows us to leverage much larger and more diverse datasets, enabling the model to learn robust features that can later be adapted to the pediatric context”.

 

Main contributions

The study offers three key contributions:

 

1. It demonstrates that pretraining AI models on large collections of adult chest X-rays improves performance when these models are fine-tuned on pediatric data.

2. It highlights the utility of lateral X-rays, which provide complementary information—especially valuable in infants and young children, where the frontal view alone may be insufficient.

3. It shows that age-specific models are preferred over models trained across all ages, due to differences in disease presentation and developmental stages among age groups.

 

Elisa López Varela, researcher at ISGlobal during the study, emphasizes: “Lateral views complement the frontal ones and are particularly valuable in infants and young children, as they help identify findings that could go unnoticed when only one projection is available.”

 

Implications for Clinical Practice and Public Health

“This solution is not intended to replace the radiologist or physician, but rather to serve as a decision-support tool: it can help prioritize cases, guide screening decisions, and facilitate early detection in resource-limited settings”, explains Begoña Santiago García, coordinator of pTBred and pediatrician at Hospital General Universitario Gregorio Marañón in Madrid. “Using lateral views and age-specific adaptation could increase diagnostic sensitivity in pediatric populations, particularly in infants and young children, where diagnosis is more complex”.

​

Transparency and Explainability

Beyond quantitative results, the team incorporated explainable AI techniques that generate visual maps highlighting the regions of the X-ray that influenced each decision (see image below). These visualizations support clinical review of the predictions and increase confidence in AI-assisted use.

​​