ptx-deconfound: trained model checkpoints for untreated-pneumothorax detection (12 configurations, 54 runs)

Trained PyTorch checkpoints for the study "Adversarial deconfounding with pseudo device labels does not remove the device shortcut in pneumothorax detection: a pre-registered, multi-seed external evaluation of twelve models". Contains one checkpoint per training run (12 model configurations, 3–5 seeds each, 54 files): the proposed U-Net/EfficientNet-B4 model with gradient-reversal device adversary and top-k mask pooling, its ablations and sensitivity variants, and re-trained baselines (U-Net ResNet34, SegFormer-B2, Group DRO, auxiliary device classifier, RAD-DINO linear probe and fine-tuned). All were trained on SIIM-ACR with the frozen splits in the code repository. Each checkpoint is the epoch selected on SIIM-ACR validation; the matching 95%-specificity decision threshold is listed in manifest.csv. Load with the code at https://github.com/yazanjer/ptx-deconfound: m = build_model(yaml.safe_load(open(config))); m.load_state_dict(torch.load(ckpt, map_location="cpu")). See MODEL_CARD.md for intended use and limitations. Not a medical device; research use only.