pre/code/predict.py

"""Standalone CBCT-only reconstruction. INPUT IS A SINGLE CBCT, NOTHING ELSE.

This entrypoint proves the system needs only a CBCT -- it loads NO label files and
contains NO reference to inst/cinst/solid/canal ground truth. The full chain is:

    CBCT (.nii.gz or DICOM folder)
      -> normalize/resample (same as training preprocess)
      -> Stage-1 (predicts tooth-core descriptors)  [image only]
      -> connected components -> per-tooth centroids -> ROIs   [image only]
      -> Stage-2 encoder -> latent -> dual SDF -> marching cubes   [image only]
      -> per-tooth nested meshes + a full-volume label NIfTI

Usage:
    python -m toothcanal.predict --config configs/default.yaml \
        --cbct /path/to/case.nii.gz --out /path/to/output_dir
    python -m toothcanal.predict --config configs/default.yaml \
        --cbct /path/to/dicom_folder --out /path/to/output_dir
"""
import os, argparse
import numpy as np
import torch
from .utils import load_config, ensure_dir, load_nii, save_nii, resample_to_spacing
from .preprocess import normalize_image
from .models import get_unet, ImplicitNet
from .infer import reconstruct_instance


def load_cbct_only(path, cfg):
    """Load a CBCT from a .nii.gz OR a DICOM directory. Returns (image, meta).
    NO label is loaded -- this function cannot see ground truth."""
    pp = cfg["preprocess"]
    if os.path.isdir(path):
        from .download import _build_image_from_dicom
        built = _build_image_from_dicom(path)
        if built is None:
            raise SystemExit(f"[predict] could not build image from DICOM dir {path}")
        img, meta = load_nii(built)
    else:
        img, meta = load_nii(path)
    img, meta = resample_to_spacing(img, meta, pp["spacing"], is_label=False)
    img = normalize_image(img, pp["clip_hu"])
    return img.astype(np.float32), meta


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--config", default="configs/default.yaml")
    ap.add_argument("--cbct", required=True, help=".nii.gz file OR a DICOM folder")
    ap.add_argument("--out", default=None, help="output directory")
    args = ap.parse_args()
    cfg = load_config(args.config)
    dev = "cuda" if torch.cuda.is_available() else "cpu"
    out_dir = ensure_dir(args.out or os.path.join(cfg["paths"]["out_dir"], "predict"))

    # 1) load CBCT only -- d holds ONLY image/spacing/origin, never any label
    img, meta = load_cbct_only(args.cbct, cfg)
    d = dict(image=img,
             spacing=np.asarray(meta["spacing"], np.float32),
             origin=np.asarray(meta.get("origin", np.zeros(3)), np.float32))
    print(f"[predict] loaded CBCT {args.cbct}  shape={img.shape}")

    # 2) Stage-1 -> predicted tooth-core instances (image only)
    stage1_ckpt = os.path.join(cfg["paths"]["out_dir"], "stage1.pt")
    if not os.path.exists(stage1_ckpt):
        raise SystemExit("[predict] missing stage1.pt -- train Stage 1 first")
    from .roi import predicted_instances
    # predicted_instances only reads d['image'] + d.get('inst') (absent here -> no match)
    inst_pred, _ = predicted_instances(d, cfg, dev, stage1_ckpt)
    d["inst"] = inst_pred
    d["cinst"] = np.zeros_like(inst_pred)        # canal comes purely from the implicit field
    ids = [int(v) for v in np.unique(inst_pred) if v > 0]
    print(f"[predict] Stage-1 found {len(ids)} teeth (from CBCT alone)")

    # 3) Stage-2 -> per-tooth nested meshes (image only)
    ckpt = torch.load(os.path.join(cfg["paths"]["out_dir"], "stage2.pt"), map_location=dev)
    n_lat = ckpt["model"]["latents.weight"].shape[0]

    class _N(ImplicitNet):
        def __init__(s): super().__init__(n_lat, cfg)
    net = _N().to(dev); net.load_state_dict(ckpt["model"]); net.eval()

    mesh_dir = ensure_dir(os.path.join(out_dir, "meshes"))
    full = np.zeros(img.shape, np.int16)
    sp = d["spacing"]; origin = d["origin"]
    for iid in ids:
        rec = reconstruct_instance(net, d, iid, cfg, dev, do_tto=False)
        if rec is None:
            continue
        off = rec["roi"]["lo_world_mm"]
        for k in ("tooth", "canal"):
            if rec[k] is not None:
                m = rec[k].copy(); m.vertices = m.vertices + off[None, :]
                m.export(os.path.join(mesh_dir, f"t{iid:02d}_{k}.stl"))
        # stamp into a full-volume label map for ITK-SNAP
        actual = rec["roi"].get("actual_size_mm", np.array([cfg["stage2"]["roi_mm"]] * 3))
        g = rec["grid"]; gsp = actual / g
        for nm, occ, val in [("tooth", rec["occ_tooth"], iid + 28),
                             ("canal", rec["occ_canal"], iid)]:
            pts = np.argwhere(occ > 0)
            if len(pts) == 0:
                continue
            world = off[None, :] + pts * gsp[None, :]
            vox = np.round((world - origin[None, :]) / sp[None, :]).astype(int)
            ok = np.all((vox >= 0) & (vox < np.array(img.shape)), axis=1)
            vox = vox[ok]
            if nm == "canal":
                full[vox[:, 0], vox[:, 1], vox[:, 2]] = val
            else:
                cur = full[vox[:, 0], vox[:, 1], vox[:, 2]]
                vv = vox[cur == 0]
                full[vv[:, 0], vv[:, 1], vv[:, 2]] = val

    save_nii(img, meta, os.path.join(out_dir, "image.nii.gz"))
    save_nii(full, meta, os.path.join(out_dir, "pred_label.nii.gz"))
    print(f"[predict] DONE. meshes in {mesh_dir}, label map at {out_dir}/pred_label.nii.gz")
    print(f"[predict] open in ITK-SNAP: main=image.nii.gz, segmentation=pred_label.nii.gz, then click Update")


if __name__ == "__main__":
    main()