| """Inference entry point for the exported CADC model repository.""" |
| from pathlib import Path |
|
|
| import numpy as np |
| import torch |
|
|
| from cadc.model import build_model |
|
|
|
|
| def load_ensemble(root, device="cpu"): |
| models = [] |
| names = [] |
| for run, arch in (("baseline", "resnet3d"), ("multiview", "multiview2d")): |
| for fold in range(5): |
| path = Path(root) / "runs" / run / f"fold{fold}" / "final.pt" |
| checkpoint = torch.load(path, map_location="cpu", weights_only=True) |
| model = build_model(arch, pretrained=False) |
| model.load_state_dict(checkpoint["model"]) |
| models.append(model.eval().to(device)) |
| names.append(f"{arch}/fold{fold}") |
| return models, names |
|
|
|
|
| @torch.inference_mode() |
| def predict_patch(models, patch, device="cpu"): |
| """Score one 64^3 HU patch, indexed (z, y, x), at 1 mm voxel spacing. |
| |
| The central 48^3 crop and HU scaling match cadc.data.NoduleDataset. |
| The returned score estimates radiologist suspicion, not cancer probability. |
| """ |
| patch = np.asarray(patch) |
| if patch.shape != (64, 64, 64): |
| raise ValueError("Expected one (64, 64, 64) nodule patch at 1 mm spacing.") |
| if patch.dtype.kind not in "iuf" or not np.isfinite(patch).all(): |
| raise ValueError("Patch must contain finite numeric Hounsfield units.") |
| if not models: |
| raise ValueError("No models loaded.") |
| crop = np.clip(patch[8:56, 8:56, 8:56].astype(np.float32), -1000, 400) |
| x = torch.from_numpy(np.ascontiguousarray((crop + 1000) / 1400))[None, None].to(device) |
| votes = [float(torch.sigmoid(model(x)).item()) for model in models] |
| return float(np.mean(votes)), votes |
|
|
