src/cell_tracking/data/wormtrack_verify/adapters.py

"""Native ASCENT and ZephIR table adapters with conservative annotation provenance."""

from __future__ import annotations

import numpy as np
import pandas as pd

XYZ = ["x_voxel", "y_voxel", "z_voxel"]
TRUSTED = {"human_corrected", "manually_verified", "verified_checkpoint"}


def _physical(df, spacing_xyz):
    for axis, scale in zip("xyz", spacing_xyz or [None] * 3, strict=True):
        df[f"{axis}_um"] = df[f"{axis}_voxel"] * scale if scale is not None else np.nan
    df["coordinate_system"] = "source_microscope"
    df["annotation_confidence"] = np.nan
    df["validity"] = True
    return df


def ascent_tracks(path, recording_id, *, predicted=False):
    """Read native ZYX CSV; physical spacing is from Zenodo 17561700 metadata."""
    df = pd.read_csv(path).rename(
        columns={
            "TrackID": "track_id",
            "ObjectID": "detection_id",
            "t": "frame_index",
            "x": "x_voxel",
            "y": "y_voxel",
            "z": "z_voxel",
        }
    )
    required = {"track_id", "detection_id", "frame_index", *XYZ}
    if not required.issubset(df.columns):
        raise ValueError(f"Missing ASCENT fields: {sorted(required - set(df.columns))}")
    df["recording_id"] = recording_id
    df["annotation_type"] = "algorithm_prediction" if predicted else "human_corrected"
    df["source_provenance"] = "ASCENT-PRED" if predicted else "ASCENT-GT-proofread-ZephIR"
    df["label_set"] = "prediction" if predicted else "reference"
    df["track_id"] = df.track_id.astype(str)
    df["detection_id"] = df.detection_id.map(lambda i: f"{'pred' if predicted else 'gt'}:{i}")
    return _physical(df, [0.243, 0.243, 1.5])


def zephir_tracks(path, metadata, recording_id):
    """Retain scorer codes; only documented ZEIR results are classified as predictions.

    Other codes remain unknown until their human/algorithm provenance is verified.
    Normalized coordinates map to voxel index x*size-0.5 using the native viewer
    convention. No voxel spacing is inferred from the neuron arrangement.
    """
    import h5py

    with h5py.File(path, "r") as h:
        df = pd.DataFrame({k: h[k][()] for k in h})
    df = df.rename(
        columns={
            "t_idx": "frame_index",
            "worldline_id": "track_id",
            "id": "detection_id",
            "provenance": "source_provenance",
        }
    )
    df["source_provenance"] = df.source_provenance.map(lambda s: s.decode("utf-8"))
    for axis in "xyz":
        df[f"{axis}_normalized"] = df[axis]
        df[f"{axis}_voxel"] = df[axis] * metadata[f"shape_{axis}"] - 0.5
        df = df.drop(columns=axis)
    df["annotation_type"] = np.where(
        df.source_provenance == "ZEIR", "algorithm_prediction", "unknown"
    )
    df["recording_id"] = recording_id
    df["label_set"] = "source_mixed"
    df["track_id"] = df.track_id.astype(str)
    df["detection_id"] = df.detection_id.map(lambda i: f"source:{i}")
    df = _physical(df, None)
    normalized = df[[f"{axis}_normalized" for axis in "xyz"]].to_numpy(dtype=float)
    df["validity"] = (
        np.isfinite(normalized).all(axis=1)
        & (normalized >= 0).all(axis=1)
        & (normalized <= 1).all(axis=1)
    )
    return df


def validate_tracks(df, *, shape_xyz=None):
    """Reject ambiguous assignments and nonfinite observed points; never clamp or interpolate."""
    required = {
        "recording_id",
        "frame_index",
        "track_id",
        "detection_id",
        *XYZ,
        "annotation_type",
        "label_set",
        "validity",
    }
    if not required.issubset(df.columns):
        raise ValueError(f"Missing fields: {sorted(required - set(df.columns))}")
    if df.empty:
        raise ValueError("Empty tracking table")
    valid = df.loc[df.validity]
    scope = ["recording_id", "label_set", "frame_index"]
    if valid.duplicated(scope + ["track_id"]).any():
        raise ValueError("Duplicate track assignments")
    if valid.duplicated(scope + ["detection_id"]).any():
        raise ValueError("Duplicate detection assignments")
    if valid.track_id.isna().any() or valid.detection_id.isna().any():
        raise ValueError("Missing identity on a valid observation")
    if not np.isfinite(valid[XYZ].to_numpy(dtype=float)).all():
        raise ValueError("Nonfinite observed coordinates")
    if (valid.frame_index < 0).any() or (valid.frame_index % 1 != 0).any():
        raise ValueError("Invalid frame index")
    # Centers use pixel boundaries [-0.5, size-0.5], including ZephIR normalized endpoints.
    if shape_xyz is not None:
        points = valid[XYZ].to_numpy(dtype=float)
        if ((points < -0.5) | (points > np.asarray(shape_xyz) - 0.5)).any():
            raise ValueError("Spatial coordinates outside source bounds")
    return {
        "observations": len(valid),
        "frames": int(valid.frame_index.nunique()),
        "tracks": int(valid.track_id.nunique()),
        "annotation_types": {
            str(k): int(v) for k, v in valid.annotation_type.value_counts().items()
        },
    }