# Automated detection of radiolucent foreign body aspiration on chest CT using deep learning

## TL;DR

A deep learning model was developed to detect hard-to-see foreign body aspiration in chest CT scans, outperforming experts and potentially reducing missed diagnoses.

## Contribution

A novel deep learning model combining airway segmentation and classification for detecting radiolucent foreign body aspiration in chest CT scans.

## Key findings

- The model achieved over 90% accuracy and balanced recall-precision metrics across three independent cohorts.
- In a blinded evaluation, the model outperformed radiologists in recall and F1 score for detecting radiolucent FBA.
- The model has potential to reduce false negatives and support clinical decision-making in high-risk cases.

## Abstract

Radiolucent foreign body aspiration (FBA) remains diagnostically challenging due to its subtle imaging signatures on chest CT scans, often leading to delayed or missed diagnoses. We present a deep learning model integrating MedpSeg, a high-precision airway segmentation method, with a convolutional classifier to detect radiolucent FBA. The model was trained and validated across three independent cohorts, demonstrating consistent performance with accuracies above 90% and balanced recall–precision metrics. In a blinded independent evaluation cohort, the model outperformed expert radiologists in both recall (71.4% vs. 35.7%) and F1 score (74.1% vs. 52.6%), highlighting its potential to reduce missed cases (false negatives) and support clinical decision-making. This study illustrates the translational potential of artificial intelligence for addressing diagnostically complex and high-risk conditions, offering an effective tool to support radiologists in the assessment of suspected radiolucent foreign body aspiration. Code is available at https://github.com/ZheChen1999/FBA_DL.