Resolution and Tiled Inference for Foreign-Object Triage in Chest Radiographs: A Validation-Calibrated Study on object-CXR

Abstract

Chest radiographs can contain external objects that obscure findings and lead to repeat imaging. We evaluate how inference configuration affects a single-class YOLO26-s detector trained at 640 px on the public object-CXR dataset. We compare whole-image inference at 640, 960, and 1,280 px with 640-px tiled inference at 20 percent overlap. Because annotations include rectangles, ellipses, and polygons, training uses enclosing boxes while evaluation scores predicted points against original shapes with free-response receiver operating characteristic (FROC); image-level performance is summarized by area under the receiver operating characteristic curve (AUC). Evaluation uses the labeled official validation split as a held-out set because the official test split is unreleased. Whole-image inference at 640 px performed best, with FROC 0.8083 and AUC 0.9362. Thresholds selected on a separate validation partition to target 95 percent image-level sensitivity yielded held-out sensitivity from 93.6 to 95.8 percent; whole-image 640-px inference flagged 32.4 clean films per 100. Four annotators achieved Fleiss kappa of 0.7432 on 86 shared objects. These findings describe one checkpoint from a 100-epoch run that stopped during epoch 43 and support conclusions only for the dataset and configurations tested.