For invisible inhaled objects, the AI stops giving a probability and points to a bronchial segment, lifting sensitivity from 18.8% to 56.3%
Synopsis
Radiolucent foreign body aspiration in adults sits at soft-tissue and mucus density on CT, so it has been found only by slice-by-slice manual reading and is easily missed. AI-PACS instead learns a standard Gaussian manifold of non-FBA airway topology from 804 unlabelled CT scans, flags deviations from that manifold as anomalies, and maps them onto the S1–S10 bronchial segments. In a 35-patient human-AI subset, sensitivity rose from 18.8% for radiologists to 56.3%, and average negative-exclusion screening time fell by about 75%.
Interpretation
Detecting radiolucent foreign body aspiration in adults shifts from a patient-level probability to a location inside the 3D airway. Earlier deep learning models were black-box classifiers that produced patient-level screening output and no precise 3D spatial coordinates for intervention. Across internal (n=268), external (n=103) and independent (n=70) cohorts, diagnostic accuracy ranged from 75.7% to 82.5%.
The system treats the geometry of the normal airway as its baseline and calls anything that departs from it anomalous. It does not depend on scarce labelled cases, pre-training instead on 804 unlabelled CT scans: 254 healthy controls from MosMedData, 250 randomized non-FBA patients from The Central Hospital of Wuhan, and 300 multi-center scans from the ATM22 challenge. During pre-training the predictive loss and the SIGReg regularizer are optimized together, and the mean L2 norm of normal airway patches declines and then plateaus.
Anomaly scores are back-projected onto the 3D airway mesh and registered to bronchial segment nomenclature, producing heatmaps for pre-operative bronchoscopy planning. A classifier only returns yes or no, leaving the interventionalist to search hundreds of slices by hand. In the 35-patient human-AI subset (16 FBA, 19 controls), AI-PACS reached 56.3% sensitivity, 78.9% specificity and 68.6% accuracy, against 18.8%, 63.2% and 42.9% for manual reading.
Perspective
The platform targets CT screening and pre-operative localization of radiolucent foreign body aspiration in adults, validated on internal, external and independent retrospective cohorts plus a 35-patient human-AI subset, and outputs heatmaps mapped to the S1–S10 segments. It is optimized for the first six to seven generations of the tracheobronchial tree, with an average end-to-end processing latency of about 60 seconds per scan.
The study is retrospective, so it cannot yet establish whether the gains in diagnostic sensitivity and screening efficiency translate into clinical endpoints such as hospital stay, complications or survival. Fragments smaller than 2 mm in the distal sub-segmental bronchi may fall below the spatial resolution of standard CT or the encoder's patch range, and high-density impacted mucus or focal endobronchial inflammatory masses can still mimic radiolucent foreign bodies in density and topology.