toothcanal/preprocess.py

"""Stage 0 preprocessing: per case, produce a compact processed bundle:
   - image  (normalized, [x,y,z] float32) at working spacing
   - sem     (0 bg / 1 tooth-solid / 2 canal)  for Stage-1 segmentation
   - inst    (tooth instance id 0..28, from tooth_solid)  for Stage-2 ROI cropping
   - cinst   (canal instance id 0..28)
   - cleaning report (json) for the QC before/after table

Run:
    python -m toothcanal.preprocess --config configs/default.yaml
"""
import os, argparse
import numpy as np
from tqdm import tqdm
from .utils import (load_config, ensure_dir, load_nii, resample_to_spacing,
                    read_json, write_json)
from .cleaning import build_instances


def normalize_image(img, clip):
    img = np.clip(img, clip[0], clip[1]).astype(np.float32)
    lo, hi = float(img.min()), float(img.max())
    if hi - lo < 1e-5:
        return np.zeros_like(img)
    return (img - lo) / (hi - lo)


def _fit_to_shape(arr, shape):
    """Crop or zero-pad a label array to an exact target shape."""
    out = np.zeros(shape, dtype=arr.dtype)
    s = [min(shape[k], arr.shape[k]) for k in range(3)]
    out[:s[0], :s[1], :s[2]] = arr[:s[0], :s[1], :s[2]]
    return out


def _resample_label_to_reference(lab, lmeta, ref_img, ref_meta):
    """Resample a label volume onto the reference image's exact voxel grid using
    SimpleITK (nearest neighbor), so label and image align even when the
    annotation was made at a different resolution."""
    import SimpleITK as sitk
    lab_t = np.transpose(lab, (2, 1, 0))
    li = sitk.GetImageFromArray(lab_t)
    li.SetSpacing(tuple(float(s) for s in lmeta["spacing"]))
    li.SetOrigin(tuple(float(o) for o in lmeta["origin"]))
    li.SetDirection(tuple(float(d) for d in lmeta["direction"]))

    ref_t = np.transpose(ref_img, (2, 1, 0))
    ri = sitk.GetImageFromArray(ref_t)
    ri.SetSpacing(tuple(float(s) for s in ref_meta["spacing"]))
    ri.SetOrigin(tuple(float(o) for o in ref_meta["origin"]))
    ri.SetDirection(tuple(float(d) for d in ref_meta["direction"]))

    out = sitk.Resample(li, ri, sitk.Transform(), sitk.sitkNearestNeighbor, 0,
                        li.GetPixelID())
    o = sitk.GetArrayFromImage(out)
    return np.transpose(o, (2, 1, 0)).astype(np.int16)


def process_case(cid, info, cfg):
    from .assemble import assemble_label
    pp = cfg["preprocess"]
    img, imeta = load_nii(info["image"])
    lab, lmeta = load_nii(info["label"])
    lab = np.rint(lab).astype(np.int16)

    # merge supplementary tooth labels for canal-only cases (e.g. 21-30)
    lab, status = assemble_label(lab, info.get("num", -1), cfg)
    if status == "canal_only":
        print(f"[preprocess] SKIP {cid}: only canal labels, no tooth labels found "
              f"(drop tooth nii into paths.extra_tooth_dir).")
        return None

    # If label grid differs from image grid (different annotation resolution),
    # first put the label onto the image's exact grid, THEN resample both together.
    if lab.shape != img.shape:
        lab = _resample_label_to_reference(lab, lmeta, img, imeta)

    # resample both to working spacing (image linear, label nearest)
    img, imeta2 = resample_to_spacing(img, imeta, pp["spacing"], is_label=False)
    lab, _ = resample_to_spacing(lab, imeta, pp["spacing"], is_label=True)
    lab = np.rint(lab).astype(np.int16)
    imeta = imeta2

    # final safety: crop/pad label to image shape so indexing never mismatches
    if lab.shape != img.shape:
        lab = _fit_to_shape(lab, img.shape)

    img = normalize_image(img, pp["clip_hu"])

    instances, report = build_instances(lab, cfg)

    sem = np.zeros(img.shape, dtype=np.uint8)      # 1 tooth-solid, 2 canal, 3 descriptor-core
    inst = np.zeros(img.shape, dtype=np.uint8)     # tooth instance id
    cinst = np.zeros(img.shape, dtype=np.uint8)    # canal instance id
    desc = np.zeros(img.shape, dtype=np.uint8)     # 1 = eroded tooth-core descriptor
    from scipy import ndimage as _ndi
    erode_iter = int(cfg["preprocess"].get("descriptor_erode_iter", 3))
    for i, d in instances.items():
        sem[d["tooth_solid"]] = 1
        inst[d["tooth_solid"]] = i
        sem[d["canal"]] = 2
        cinst[d["canal"]] = i
        # descriptor = tooth body eroded so neighboring teeth separate into cores
        core = _ndi.binary_erosion(d["tooth_solid"], iterations=erode_iter)
        if not core.any():                          # tiny tooth: keep a seed voxel
            core = d["tooth_solid"]
        desc[core] = i                              # store instance id in the core map
    # Stage-1 semantic target: 0 bg / 1 tooth / 2 canal / 3 descriptor-core
    sem[desc > 0] = 3

    out_dir = ensure_dir(cfg["paths"]["proc_dir"])
    np.savez_compressed(
        os.path.join(out_dir, f"{cid}.npz"),
        image=img.astype(np.float32),
        sem=sem, inst=inst, cinst=cinst, desc=desc,
        spacing=np.array(imeta["spacing"], dtype=np.float32),
        origin=np.array(imeta["origin"], dtype=np.float32),
        n_instances=len(instances),
    )
    return dict(case=cid, n_instances=len(instances), cleaning=report,
                status=status, shape=list(img.shape))


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--config", default="configs/default.yaml")
    ap.add_argument("--limit", type=int, default=0, help="process only first N (debug)")
    ap.add_argument("--force", action="store_true",
                    help="reprocess cases even if their .npz already exists")
    args = ap.parse_args()
    cfg = load_config(args.config)

    manifest = read_json(os.path.join(cfg["paths"]["raw_dir"], "manifest.json"))
    items = list(manifest.items())
    if args.limit:
        items = items[:args.limit]

    ensure_dir(cfg["paths"]["proc_dir"])
    if not args.force:
        before = len(items)
        items = [(cid, info) for cid, info in items
                 if not os.path.exists(os.path.join(cfg["paths"]["proc_dir"], f"{cid}.npz"))]
        skipped = before - len(items)
        if skipped:
            print(f"[preprocess] skipping {skipped} already-processed case(s); "
                  f"will process {len(items)}.")
    summary = []
    for cid, info in tqdm(items, desc="preprocess"):
        try:
            r = process_case(cid, info, cfg)
            if r is not None:
                summary.append(r)
        except Exception as e:
            print(f"[preprocess] FAILED {cid}: {e}")
    write_json(summary, os.path.join(cfg["paths"]["proc_dir"], "preprocess_report.json"))

    # ---- aggregate report (handles both 'offset' per-instance and 'geometric' summary) ----
    nb = na = vb = va = specks = 0
    reassigned = orphaned = ncomp = geo_cases = 0
    for s in summary:
        for c in s["cleaning"]:
            if "pairing" in c:                     # geometric re-pairing summary
                reassigned += c.get("voxels_reassigned", 0)
                orphaned += c.get("voxels_orphaned", 0)
                ncomp += c.get("n_components", 0)
                geo_cases += 1
            else:                                  # legacy offset per-instance cleaning
                nb += c.get("n_components_before", 0); na += c.get("n_components_after", 0)
                vb += c.get("voxels_before", 0); va += c.get("voxels_after", 0)
                specks += c.get("removed_specks", 0)
    if geo_cases:
        print("\n========== canal<->tooth geometric re-pairing ==========")
        print(f"cases re-paired: {geo_cases} | canal components: {ncomp}")
        print(f"voxels re-assigned to correct tooth: {reassigned} | orphaned/dropped: {orphaned}")
        print("========================================================")
    if vb:
        print("\n========== canal cleaning (before / after) ==========")
        print(f"components: {nb}  ->  {na}   (removed {specks} specks)")
        print(f"voxels retained: {100.0 * va / vb:.2f}%  (lost {100.0 * (vb - va) / vb:.2f}%)")
        print("=====================================================")


if __name__ == "__main__":
    main()