A research team from the University of Southampton has developed an AI-powered diagnostic tool that can identify hard-to-detect objects lodged in the airways, often more accurately than experienced radiologists. The innovation could significantly improve early diagnosis and treatment of foreign body aspiration (FBA), a condition that can lead to coughing, choking, and even life-threatening respiratory complications.
The study, Automated Detection of Radiolucent Foreign Body Aspiration on Chest CT Using Deep Learning, demonstrates how artificial intelligence can outperform human experts in complex imaging interpretation, particularly when subtle findings are nearly invisible on standard scans.
Detecting what the human eye can’t see
Foreign body aspiration occurs when a small object, such as food, plant matter, or shell fragments, becomes lodged in a patient’s airway. Detecting these radiolucent materials is notoriously difficult because they are invisible on X-rays and only faintly visible on CT scans. As a result, up to three-quarters of adult FBA cases may initially go undiagnosed or be misinterpreted, delaying critical care.
“These objects can be incredibly subtle, even for seasoned radiologists,” explains Zhe Chen, Ph.D. researcher and co-first author of the study. “Our AI system acts as a second pair of eyes, helping clinicians spot potential obstructions earlier and with greater confidence.” The study was published in npj Digital Medicine.
Developed by Dr. Yihua Wang, Dr. Zehor Belkhatir, and Prof. Rob Ewing in collaboration with researchers from Wuhan, China, the AI tool integrates deep learning with an advanced airway mapping algorithm known as MedpSeg. This combination enables the model to detect minute patterns in CT scans that may indicate the presence of a hidden foreign object.
Outperforming human experts
The model was trained and validated using data from over 400 patients, sourced from three independent hospital cohorts in China. It was then benchmarked against the performance of three senior radiologists, each with more than ten years of diagnostic experience.
The test set included 70 CT scans, 14 of which had confirmed cases of radiolucent FBA verified through bronchoscopy. While radiologists demonstrated perfect precision, meaning no false positives, they detected only 36% of true cases, underscoring the difficulty of identifying these subtle obstructions.
The AI model, by contrast, detected 71% of true cases, achieving an F1 score of 74%, compared to 53% for human readers. Although the AI produced some false positives, it dramatically improved overall detection rates, ensuring fewer dangerous obstructions go unnoticed.
“The results highlight how AI can enhance diagnostic accuracy in situations where traditional imaging interpretation reaches its limits,” says Dr. Wang, lead author of the study. “By providing an additional layer of analysis, AI systems can help clinicians make faster and more informed decisions.”
Supporting clinicians
The researchers emphasize that the AI system is designed as a clinical support tool, not a replacement for radiologists. Its role is to flag potential abnormalities and assist human experts in reviewing complex or ambiguous cases. “This kind of assistive technology doesn’t take over the radiologist’s role, it strengthens it,” notes Prof. Ewing. “By combining human expertise with AI precision, we can minimize missed diagnoses and improve patient safety.”
The Southampton team now plans to expand testing through multi-center clinical trials involving larger and more diverse patient populations. This will help refine the model, reduce bias, and validate its performance across various imaging systems and clinical settings.
The ultimate goal is to integrate the AI system into hospital workflows, enabling real-time analysis of chest CT scans and supporting faster, more consistent detection of airway obstructions. “AI will increasingly become a trusted partner in medical imaging,” concludes Dr. Wang. “By augmenting human expertise, it has the power to transform diagnosis and ensure that no subtle, life-threatening condition goes unnoticed.”
AI in radiology
Researchers at the Australian e-Health Research Centre (AEHRC) are working with hospitals to develop AI models that help radiologists analyse X-ray images, with the aim of reducing their workload and speeding up diagnoses. The latest generation uses visual language models (VLMs), which combine image recognition with natural language processing. These systems can not only “see” X-rays, but also automatically generate medical reports, for example for heart and lung examinations.
By using thousands of existing X-ray images and reports as training data, the AI learns to write accurate radiological reports independently. Integration of emergency department data, such as vital signs and medication, further improved performance. Although still in the research phase, the results show promising potential. AEHRC emphasises that AI is intended as support, not as a replacement for medical specialists.
