Researchers at the University of Southampton have developed an AI-tool capable of detecting hard-to-see foreign objects lodged in patients’ airways and found it can outperform experienced radiologists in certain challenging scenarios.
The study, published in npj Digital Medicine, tackled the problem of radiolucent foreign-body aspiration (FBA) — when objects inhaled into the airway are invisible or faint on CT scans and X-rays. These can cause coughing, breathing-difficulty and serious complications if missed. The researchers trained a deep-learning model that combined an airway-mapping algorithm (“MedpSeg”) with a neural network to analyse CT images, using data from over 400 patients from three independent groups.
In head-to-head comparison, three radiologists (each with over 10 years of experience) reviewed 70 CT scans (14 confirmed FBA cases) and achieved a detection rate of only 36 %, despite perfect precision when they did identify a case. The AI model detected 71 % of the FBA cases (vs. 36 % by the radiologists) and achieved a higher F1 score (74 % vs. 53 %). However, the model had a lower precision (77 %) compared to the radiologists’ perfect precision (100 %).
The findings suggest AI tools like this could serve as a “second set of eyes” in radiology — flagging subtle cases that might be overlooked and helping prompt earlier, more accurate diagnosis and intervention. The authors caution, however, that while promising, such tools should complement human expertise rather than replace it, as trade-offs like false positives still remain and clinical integration will require further validation and workflow design.
