david vinu
Looking for my twin?

Building Foreign Object Segmentation datasets for Chest X-Rays with Synthetic-trained proposals

Constantin Seibold, David Vinu, Matthias A. Fink

ECCV 2026 DCA-MI·26 August 2026

CheXtra trains a class-agnostic proposer on synthetic composited foreign objects to localize devices and foreign objects in real chest X-rays while a companion engine (Chevron) allows users to categorize these proposals on a cluster level.

Abstract

Foreign objects and support devices such as catheters, lines, retained surgical items, and ingested objects are among the most common abnormal findings on chest radiographs, and their misplacement is frequent and time-critical. Support lines are malpositioned in 3-14% of cases. Pixel-level masks for these objects barely exist, and every cheap label-free route (i.e. foundation segmenters, unsupervised discovery, ...) breaks on thin, rare, or unmentioned devices. We train CheXtra, a class-agnostic object-proposal generator that locates foreign-object regions for a human to verify and refine. We composite procedurally rendered foreign objects onto real, device-free radiographs, which yields exact masks for diverse appearances. Against thirteen off-the-shelf, unsupervised, and cross-domain proposers on two unseen real chest-X-ray datasets, CheXtra is the only label-free method working on both domains, at Average Recall 0.50 and 0.58. The strongest baseline reaches 0.72 on one domain and falls to 0.05 on the other. At the few-proposal budget a curator uses, CheXtra recalls thin catheters at 0.44, against 0.01 or below for every off-the-shelf method. Ablations identify device-appearance diversity as the dominant transfer factor, with no consistent benefit from anatomical placement realism and limited impact of corpus scaling. A human-in-the-loop engine, Chevron, allows reassembling of fragmentations into masks and class-assignment of proposals by clustering, retrieval and manual curation. We release the code at github.com/ConstantinSeibold/CXR_SynthFB.

keywords
Synthetic data, Chest X-ray, Foreign objects, Object proposals, Data curation
links
openreview↗code↗
licence
CC BY 4.0↗

cite

@inproceedings{seibold2026building,
  title     = {Building Foreign Object Segmentation datasets for Chest X-Rays with Synthetic-trained proposals},
  author    = {Seibold, Constantin and Vinu, David and Fink, Matthias A.},
  booktitle = {ECCV 2026 Workshop DCA-MI},
  year      = {2026},
  url       = {https://openreview.net/forum?id=RoWwewngwy}
}