Thoracic X-ray: CoRE teaches AI that the same anomaly does not say the same thing depending on the lung location

A model that links each anomaly to its anatomical region and the lesions that often accompany it better locates signs on chest radios, including on unknown bases.

To read a chest X-ray, a radiologist takes into account both the location of an abnormality and the other signs that accompany it. However, most AI models reason on global labels ("opacity", "effusion") without distinguishing the lobe or region concerned. Two Korean researchers present on arXiv CoRE (Co-occurrence-guided Region Evidence), which associates each abnormality with its anatomical region.

The system groups together the region-anomaly pairs that often appear together and uses them to guide the learning of a frozen Vision Transformer, adapted by "visual prompt tuning"; a second branch, supervised region by region, corrects the predictions. Tested on MIMIC-CXR with Chest ImaGenome annotations, CoRE classifies anomalies better by region than reference approaches, locates them more precisely on MS-CXR and also improves detection on two external databases, VinDr-CXR and CheXpert, without prior adaptation.