HalmosiL/nodule-segresnet-3d-wide-sw

3D nodule segmentation — SegResNet (wide), sliding-window recipe

Two-stage 3D nodule segmentation on lung-cropped CT at native 1 mm resolution. Trained on 128^3 nodule-biased patches (Focal Tversky + weighted CE, nodule CE weight 10), inference via 128^3 sliding windows (overlap 0.5, Gaussian blending).

Test split of the released split.csv (308 series): nodule Dice (micro) 0.6662, precision 0.5448, recall 0.8571; instance recall 0.8697, instance precision 0.1125.

Files: model.pth (weights-only state dict) + config.yaml (full training configuration). Training and evaluation code, lung bounding boxes and reports: szegedai/nodule-segmentation-baselines.