Finding localizer

Measured on held-out scans the model never saw during training. The first run reported 92.4%, and that figure was memory rather than quality — 204 of the 250 test positives were in its own training set. The split is now made by patient with the same seed the training notebook uses, and both scripts read one shared list so the comparison cannot drift.

Models that do this already exist — ChEX, BioViL, Foundation X — and every one of them was unusable here for the same reason, which is the licence rather than the quality. BioViL is published under MIT and its own model card still says that any deployed use, commercial or otherwise, is out of scope. ChEX ships weights but was trained on MIMIC-CXR, MS-CXR and VinDr-CXR, all of which are PhysioNet credentialed data, research and education only.

The RSNA Pneumonia Detection Challenge is the one chest X-ray dataset with drawn regions that permits commercial use with attribution. That makes the chain clean end to end: our code, permitted data, our weights. The encoder starts from ImageNet rather than from our own PadChest classifier for the same reason — on held-out data the two were indistinguishable, so the tie was broken on provenance.

The highlighted area is where the model responded most strongly. Its edges are that response falling off, not a measured boundary of anything, and its shape says nothing about the shape of what is there. It runs at 512×512 rather than the classifiers' 224 precisely so that the map is not a coarse grid stretched over a chest — the resolution is the feature.

Radiological Society of North America. RSNA Pneumonia Detection Challenge (2018). Annotations by the RSNA and the Society of Thoracic Radiology.