File size: 21,272 Bytes
3670eb8
 
 
 
 
 
 
 
 
 
 
 
 
837970b
7215c7f
3670eb8
5833bcb
3670eb8
c68ac12
5833bcb
3670eb8
0af6558
3670eb8
0af6558
837970b
3670eb8
 
 
 
0af6558
 
3670eb8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5833bcb
 
 
 
 
 
 
 
 
 
 
 
 
782869f
 
 
 
 
 
 
 
 
5833bcb
 
 
 
 
 
 
 
 
 
 
 
782869f
 
 
 
 
 
 
 
 
853cb9e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4d00e0f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
853cb9e
4d00e0f
853cb9e
4d00e0f
 
853cb9e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4d00e0f
 
 
 
 
 
 
853cb9e
 
 
 
 
 
 
 
4d00e0f
 
 
853cb9e
4d00e0f
 
 
 
 
853cb9e
 
 
 
 
 
 
 
c68ac12
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f7e54c4
 
 
 
 
 
 
c68ac12
 
 
 
 
 
 
 
 
 
f7e54c4
 
 
 
 
 
 
c68ac12
 
 
 
 
 
 
 
 
 
7215c7f
 
 
 
 
 
 
b18a079
 
 
 
 
 
 
 
7215c7f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
837970b
7215c7f
 
 
 
 
 
 
 
 
 
 
 
 
 
b18a079
7215c7f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
837970b
7215c7f
 
b18a079
837970b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b18a079
7215c7f
837970b
 
 
 
b18a079
7215c7f
837970b
 
0af6558
7215c7f
 
 
 
 
837970b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7215c7f
1068ed1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7215c7f
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
"""MRI (magnetic resonance imaging) pipeline.

Loads NIfTI volumes (`.nii` / `.nii.gz`), applies a brain mask, harmonizes
across sites with ComBat (`neuroHarmonize`), and writes per-subject ROI
statistics as a model-ready Parquet at `data/processed/mri_features.parquet`.

Follows the Data Readiness contract in AGENTS.md §4 and the Parquet storage
convention in §6: schema validity, domain validity (drop NaN/inf volumes
with a logged WARNING), determinism (ComBat is RNG-free given fixed input),
traceability (in/out/dropped counts at INFO), and idempotent overwrite.
"""
from __future__ import annotations

import time
from pathlib import Path

import nibabel as nib
import numpy as np
import pandas as pd
from scipy import ndimage as scipy_ndimage

from src.core.determinism import pin_threads
from src.core.logger import get_logger
from src.core.storage import write_parquet
from src.core.tracking import track_pipeline_run

logger = get_logger(__name__)

# Pin BLAS / OpenMP / pyarrow to single-threaded mode so byte-determinism
# (AGENTS.md §4 rule 3) holds across hardware. See src.core.determinism.
pin_threads()


def is_valid_volume(volume: np.ndarray | None) -> bool:
    """Return True iff `volume` is a non-empty 3-D numeric array with no NaN/inf.

    Used to drop corrupted volumes before masking + feature extraction.
    Defensive against the full set of garbage we expect from real archives:
    lists, None, NaN/inf samples, zero-sized arrays, string-dtype arrays.
    """
    if not isinstance(volume, np.ndarray):
        return False
    if volume.ndim != 3:
        return False
    if volume.size == 0:
        return False
    if not np.issubdtype(volume.dtype, np.number):
        return False
    if not np.all(np.isfinite(volume)):
        return False
    return True


def mask_brain(
    volume: np.ndarray,
    intensity_threshold: float | None = None,
) -> np.ndarray:
    """Build a brain mask from a 3-D MRI volume.

    Two-step pipeline:
      1. Intensity threshold: keep voxels above `intensity_threshold`. When
         `None`, use the volume's mean as a robust auto-threshold (works on
         the synthetic fixture where brain ≫ background; for real data the
         caller should pass an Otsu or BET-derived threshold explicitly).
      2. Morphological opening (`scipy.ndimage.binary_opening`, 6-connectivity,
         iterations=1) to remove isolated noise voxels and disconnected
         fragments. Note: thin features (< 3 voxels wide along any axis pair)
         may be eroded entirely; for production data with cortical sheets or
         sulcal bridges, prefer 26-connectivity or pass `iterations=0` upstream.

    If the resulting mask is all-False (e.g. caller passed a threshold above
    the volume's max intensity, or the volume is constant-valued), a WARNING
    is emitted so silent feature-zeroing is visible in production logs.

    Args:
        volume: 3-D numeric `np.ndarray` (must satisfy `is_valid_volume`).
        intensity_threshold: Voxel-intensity floor. `None` → use `volume.mean()`.

    Returns:
        A boolean `np.ndarray` of the same shape as `volume`. True = brain.
    """
    if intensity_threshold is None:
        intensity_threshold = float(volume.mean())

    raw = volume > intensity_threshold
    cleaned = scipy_ndimage.binary_opening(raw, iterations=1).astype(bool)
    if not cleaned.any():
        logger.warning(
            "mask_brain produced an all-False mask "
            "(volume min=%.4f, max=%.4f, threshold=%.4f); "
            "downstream features for this volume will be all-zero.",
            float(volume.min()), float(volume.max()), intensity_threshold,
        )
    return cleaned


# Default ROI partition: split a (D, H, W) volume into 2×2×2 = 8 octant ROIs.
# Octant index follows binary (z, y, x) ordering: 0..7.
DEFAULT_N_ROI_AXES: tuple[int, int, int] = (2, 2, 2)


def _roi_slices(
    shape: tuple[int, int, int],
    n_roi_axes: tuple[int, int, int],
) -> list[tuple[slice, slice, slice]]:
    """Generate the ROI slice list in deterministic (z, y, x) octant order."""
    nz, ny, nx = n_roi_axes
    dz, dy, dx = shape
    bins_z = np.array_split(np.arange(dz), nz)
    bins_y = np.array_split(np.arange(dy), ny)
    bins_x = np.array_split(np.arange(dx), nx)
    out: list[tuple[slice, slice, slice]] = []
    for bz in bins_z:
        for by in bins_y:
            for bx in bins_x:
                out.append((
                    slice(bz[0], bz[-1] + 1),
                    slice(by[0], by[-1] + 1),
                    slice(bx[0], bx[-1] + 1),
                ))
    return out


# Statistical functions, bound to their column-label names. The `ROI_STATS`
# tuple below is derived from this list so labels and computations cannot
# drift out of sync (a class of bug the prior parallel-list design was
# vulnerable to — same pattern as EEG's _STATS_FUNCS).
#
# `mean`/`std` use NumPy with `ddof=0` (biased / population estimators).
# `p10`/`p50`/`p90` use `np.percentile` default linear interpolation.
# `voxel_count` is stored as float for column-uniformity in the eventual
# Parquet, but always represents a whole number (assertable via
# `v == float(int(v))`).
_ROI_STATS_FUNCS: tuple[tuple[str, "object"], ...] = (
    ("mean", lambda v: float(v.mean())),
    ("std", lambda v: float(v.std())),
    ("p10", lambda v: float(np.percentile(v, 10))),
    ("p50", lambda v: float(np.percentile(v, 50))),
    ("p90", lambda v: float(np.percentile(v, 90))),
    ("voxel_count", lambda v: float(v.size)),
)
ROI_STATS: tuple[str, ...] = tuple(name for name, _ in _ROI_STATS_FUNCS)


def _roi_stats_for(values: np.ndarray) -> dict[str, float]:
    """Compute the ROI stats. Empty array → all 0.0 (no-NaN contract)."""
    if values.size == 0:
        return {name: 0.0 for name, _ in _ROI_STATS_FUNCS}
    return {name: fn(values) for name, fn in _ROI_STATS_FUNCS}


def extract_features_from_volume(
    volume: np.ndarray,
    mask: np.ndarray,
    n_roi_axes: tuple[int, int, int] = DEFAULT_N_ROI_AXES,
) -> dict[str, float]:
    """Compute per-ROI summary statistics from a masked volume.

    The volume is partitioned into ``prod(n_roi_axes)`` axis-aligned octants
    in deterministic (z, y, x) order. For each ROI, intensity values from
    voxels where `mask` is True are summarized via mean / std / 10th, 50th,
    90th percentile / voxel count. Empty ROIs (no mask voxels) report all
    zeros so the resulting Parquet has no NaN values.

    Statistical conventions:
      - ``mean`` / ``std`` use ``ddof=0`` (biased / population estimators).
      - ``p10`` / ``p50`` / ``p90`` use ``np.percentile`` with the default
        linear interpolation.
      - ``voxel_count`` is stored as float for column uniformity but always
        represents a whole number.

    Args:
        volume: 3-D numeric `np.ndarray` (already validated).
        mask: Boolean `np.ndarray` of the same shape (from `mask_brain`).
        n_roi_axes: ROI grid along (z, y, x). Default `(2, 2, 2)` → 8 ROIs.

    Returns:
        Flat dict `{"feat_roi{i}_{stat}": float}` of length
        ``prod(n_roi_axes) * len(ROI_STATS)``.

    Raises:
        ValueError: if `volume.shape` and `mask.shape` differ.
    """
    if volume.shape != mask.shape:
        raise ValueError(
            f"volume.shape {volume.shape} != mask.shape {mask.shape}"
        )

    feats: dict[str, float] = {}
    slices = _roi_slices(volume.shape, n_roi_axes)
    for i, sl in enumerate(slices):
        roi_values = volume[sl][mask[sl]]
        stats = _roi_stats_for(roi_values)
        for stat_name, stat_val in stats.items():
            feats[f"feat_roi{i}_{stat_name}"] = stat_val
    return feats


def harmonize_combat(
    features: pd.DataFrame,
    sites: pd.Series,
    feature_cols: list[str],
) -> pd.DataFrame:
    """Apply ComBat harmonization across sites to remove site-level domain shift.

    Wraps `neuroHarmonize.harmonizationLearn` which fits a parametric ComBat
    model (no internal RNG → byte-deterministic given fixed input). Only
    `feature_cols` are harmonized; other columns in `features` (e.g.
    metadata) are not touched by this function — callers should join after.

    Args:
        features: DataFrame with at least the columns listed in `feature_cols`.
        sites: Site label per row (length must match `len(features)`).
        feature_cols: Names of the columns to harmonize.

    Returns:
        A new DataFrame of identical shape & column order to
        `features[feature_cols]`, with ComBat-harmonized values.

    Raises:
        ValueError: if fewer than 2 distinct sites are present.
    """
    from neuroHarmonize import harmonizationLearn

    if not feature_cols:
        raise ValueError("feature_cols must be a non-empty list")
    if len(features) != len(sites):
        raise ValueError(
            f"features has {len(features)} rows but sites has {len(sites)} elements"
        )

    if sites.nunique() < 2:
        raise ValueError(
            f"ComBat requires at least 2 sites; got {sites.nunique()} "
            f"({sites.unique().tolist()})"
        )

    matrix = features[feature_cols].to_numpy(dtype=np.float64)
    covars = pd.DataFrame({"SITE": sites.to_numpy()})

    _, harmonized = harmonizationLearn(matrix, covars)
    # Defensive: with OMP/OPENBLAS/MKL_NUM_THREADS=1 (set at module import,
    # per AGENTS.md §4), harmonizationLearn is already bit-identical across
    # calls. np.round(14) provides an additional determinism boundary for
    # environments where those env pins are overridden before module load
    # (e.g. a sub-process that re-exports a thread count). It discards ~5
    # trailing-mantissa bits, which is well below ComBat's biological
    # effect-size precision floor.
    out = pd.DataFrame(
        np.round(np.asarray(harmonized, dtype=np.float64), 14),
        columns=list(feature_cols),
        index=features.index,
    )
    logger.info(
        "ComBat harmonized %d rows × %d features across %d sites",
        len(out), len(feature_cols), sites.nunique(),
    )
    return out


# Default I/O paths for the MRI pipeline. Override via run_pipeline() args.
DEFAULT_INPUT = Path("data/raw/mri")
DEFAULT_OUTPUT = Path("data/processed/mri_features.parquet")


# Variance floor used to decide whether a feature column is "constant" for
# ComBat. Strict ``std() > 0`` would still send near-zero-variance columns
# (e.g. ULP-level differences) into ComBat, where var_pooled ≈ 0 produces
# NaN. 1e-8 is well above machine epsilon and far below any biologically
# meaningful signal variance.
_MIN_VAR_THRESHOLD: float = 1e-8


def _list_nifti_volumes(input_dir: Path) -> list[Path]:
    """Return sorted list of .nii / .nii.gz files in `input_dir`."""
    return sorted(
        p for p in input_dir.iterdir()
        if p.suffix == ".nii" or p.name.endswith(".nii.gz")
    )


def run_pipeline(
    input_dir: Path = DEFAULT_INPUT,
    sites_csv: Path | None = None,
    output_path: Path = DEFAULT_OUTPUT,
    intensity_threshold: float | None = None,
    n_roi_axes: tuple[int, int, int] = DEFAULT_N_ROI_AXES,
) -> None:
    """Run the MRI pipeline end-to-end: NIfTI directory → harmonized Parquet.

    For each `subject_id.nii(.gz)` in `input_dir`, validates the volume,
    masks the brain, computes per-ROI statistics, then harmonizes across
    sites (column "site" of `sites_csv`, joined on "subject_id") via ComBat.
    Output is float64 Parquet at `output_path`.

    Args:
        input_dir: Directory containing one NIfTI per subject and a
            `sites.csv` (or `sites_csv` override) with columns
            `subject_id, site`.
        sites_csv: Path to the site-covariates CSV. If `None`, defaults to
            `input_dir / "sites.csv"`.
        output_path: Where to write the processed feature Parquet file.
        intensity_threshold: Brain-mask intensity floor. `None` → per-volume
            mean (see `mask_brain`).
        n_roi_axes: ROI grid (z, y, x).

    Raises:
        FileNotFoundError: if `input_dir` does not exist.
        IsADirectoryError: if `output_path` resolves to an existing directory.
        KeyError: if `sites_csv` is missing a site for some subject.
    """
    input_dir = Path(input_dir)
    output_path = Path(output_path)
    if not input_dir.exists():
        raise FileNotFoundError(f"MRI input directory not found: {input_dir}")
    sites_csv = Path(sites_csv) if sites_csv is not None else input_dir / "sites.csv"
    if not sites_csv.exists():
        raise FileNotFoundError(f"sites_csv not found: {sites_csv}")

    started = time.perf_counter()
    logger.info("Reading MRI volumes from %s", input_dir)
    nifti_paths = _list_nifti_volumes(input_dir)
    sites_df = pd.read_csv(sites_csv)

    rows: list[dict[str, float | str]] = []
    invalid_subject_ids: list[str] = []
    for path in nifti_paths:
        subject_id = path.name.removesuffix(".nii.gz").removesuffix(".nii")
        volume = nib.load(path).get_fdata()
        if not is_valid_volume(volume):
            invalid_subject_ids.append(subject_id)
            continue
        mask = mask_brain(volume, intensity_threshold=intensity_threshold)
        feats = extract_features_from_volume(volume, mask, n_roi_axes=n_roi_axes)
        rows.append({"subject_id": subject_id, **feats})

    n_total = len(nifti_paths)
    n_dropped = len(invalid_subject_ids)
    if n_dropped:
        display = invalid_subject_ids[:10]
        suffix = (
            f"... (+{n_dropped - 10} more)" if n_dropped > 10 else ""
        )
        logger.warning(
            "Dropping %d/%d volumes with invalid samples (subjects=%s%s)",
            n_dropped, n_total, display, suffix,
        )

    feature_cols = [
        f"feat_roi{i}_{stat}"
        for i in range(int(np.prod(n_roi_axes)))
        for stat in ROI_STATS
    ]

    if not rows:
        logger.info(
            "Feature extraction complete: in=%d, out=0, dropped=%d (%.2f%%)",
            n_total, n_dropped, 100.0 * n_dropped / max(n_total, 1),
        )
        final = pd.DataFrame(
            columns=["subject_id", "site", *feature_cols]
        ).astype({c: np.float64 for c in feature_cols})
    else:
        raw_features = pd.DataFrame(rows)
        raw_features = raw_features.merge(sites_df, on="subject_id", how="left")
        if raw_features["site"].isna().any():
            missing = raw_features.loc[raw_features["site"].isna(), "subject_id"].tolist()
            raise KeyError(
                f"sites_csv missing site assignment for subjects: {missing}"
            )

        # ComBat cannot handle (near-)zero-variance columns: var_pooled ≈ 0 produces
        # NaN. Split feature_cols on a strictly-positive variance floor so ULP-level
        # noise is treated as constant.
        col_std = raw_features[feature_cols].std()
        var_feature_cols = [c for c in feature_cols if col_std[c] > _MIN_VAR_THRESHOLD]
        zero_var_cols = [c for c in feature_cols if col_std[c] <= _MIN_VAR_THRESHOLD]

        if not var_feature_cols:
            # Degenerate dataset: every feature is essentially constant. ComBat has
            # no signal to harmonize on; pass all columns through and warn.
            logger.warning(
                "All %d feature columns have variance ≤ %.1e; ComBat skipped "
                "(output contains unharmonized features).",
                len(feature_cols), _MIN_VAR_THRESHOLD,
            )
            harmonized = raw_features[feature_cols].copy()
        else:
            harmonized = harmonize_combat(
                raw_features, raw_features["site"], var_feature_cols,
            )
            # Re-attach zero-variance columns (unchanged) and restore the original
            # column order.
            for c in zero_var_cols:
                harmonized[c] = raw_features[c].to_numpy()
            harmonized = harmonized[feature_cols]

        final = pd.concat(
            [raw_features[["subject_id", "site"]].reset_index(drop=True),
             harmonized.reset_index(drop=True)],
            axis=1,
        )

        logger.info(
            "Feature extraction complete: in=%d, out=%d, dropped=%d (%.2f%%)",
            n_total, len(final), n_dropped, 100.0 * n_dropped / max(n_total, 1),
        )

    # Parquet preserves dtypes (float64 features stay float64) and is
    # byte-deterministic with single-threaded snappy. AGENTS.md §6. Unconditional
    # so the §4-rule-4 traceability log fires for both empty and non-empty paths.
    write_parquet(final, output_path)
    logger.info(
        "Wrote processed features to %s (rows=%d, cols=%d)",
        output_path, len(final), final.shape[1],
    )

    duration_sec = time.perf_counter() - started

    with track_pipeline_run(
        experiment_name="mri_pipeline",
        params={
            "input_dir": str(input_dir),
            "sites_csv": str(sites_csv),
            "output_path": str(output_path),
            "intensity_threshold": str(intensity_threshold),
            "n_roi_axes": str(n_roi_axes),
        },
        metrics={
            "subjects_in": float(n_total),
            "subjects_out": float(len(final)),
            "subjects_dropped": float(n_dropped),
            "duration_sec": duration_sec,
        },
        artifact_path=output_path,
    ):
        pass


def compute_harmonization_diagnostics(
    input_dir: Path,
    sites_csv: Path | None = None,
    intensity_threshold: float | None = None,
    n_roi_axes: tuple[int, int, int] = DEFAULT_N_ROI_AXES,
) -> pd.DataFrame:
    """Run the MRI pipeline twice — pre-ComBat features and post-ComBat —
    and return a long-format DataFrame ready for visualization.

    Output columns: ``subject_id``, ``site``, ``feature``, ``feature_value``,
    ``harmonization_state`` ('Pre-ComBat' or 'Post-ComBat').

    Used by the FastAPI ``/pipeline/mri/diagnostics`` endpoint to feed the
    Streamlit MRI tab's KDE / histogram comparison plot.

    Raises:
        FileNotFoundError: if ``input_dir`` does not exist.
        KeyError: if any subject is missing a site assignment.
    """
    input_dir = Path(input_dir)
    if not input_dir.exists():
        raise FileNotFoundError(f"MRI input directory not found: {input_dir}")
    sites_csv = Path(sites_csv) if sites_csv is not None else input_dir / "sites.csv"
    sites_df = pd.read_csv(sites_csv)

    feature_cols = [
        f"feat_roi{i}_{stat}"
        for i in range(int(np.prod(n_roi_axes)))
        for stat in ROI_STATS
    ]

    rows: list[dict[str, object]] = []
    for nifti_path in sorted(input_dir.glob("*.nii*")):
        subject_id = nifti_path.stem.replace(".nii", "")
        volume = nib.load(nifti_path).get_fdata()
        if not is_valid_volume(volume):
            continue
        mask = mask_brain(volume, intensity_threshold=intensity_threshold)
        feats = extract_features_from_volume(
            volume, mask, n_roi_axes=n_roi_axes,
        )
        row: dict[str, object] = {"subject_id": subject_id}
        row.update(feats)
        rows.append(row)

    if not rows:
        return pd.DataFrame(columns=[
            "subject_id", "site", "feature", "feature_value", "harmonization_state",
        ])

    raw_features = pd.DataFrame(rows).merge(sites_df, on="subject_id", how="left")
    if raw_features["site"].isna().any():
        missing = raw_features.loc[raw_features["site"].isna(), "subject_id"].tolist()
        raise KeyError(
            f"sites_csv missing site assignment for subjects: {missing}"
        )

    # Post-ComBat: variance-aware harmonization. Reuses the same logic as
    # run_pipeline so diagnostics reflect production behavior exactly.
    col_std = raw_features[feature_cols].std()
    var_feature_cols = [
        c for c in feature_cols if col_std[c] > _MIN_VAR_THRESHOLD
    ]
    zero_var_cols = [
        c for c in feature_cols if col_std[c] <= _MIN_VAR_THRESHOLD
    ]
    if not var_feature_cols:
        harmonized = raw_features[feature_cols].copy()
    else:
        harmonized = harmonize_combat(
            raw_features, raw_features["site"], var_feature_cols,
        )
        for c in zero_var_cols:
            harmonized[c] = raw_features[c].to_numpy()
        harmonized = harmonized[feature_cols]
    post_features = pd.concat(
        [raw_features[["subject_id", "site"]].reset_index(drop=True),
         harmonized.reset_index(drop=True)],
        axis=1,
    )

    long_pre = raw_features.melt(
        id_vars=["subject_id", "site"], value_vars=feature_cols,
        var_name="feature", value_name="feature_value",
    )
    long_pre["harmonization_state"] = "Pre-ComBat"
    long_post = post_features.melt(
        id_vars=["subject_id", "site"], value_vars=feature_cols,
        var_name="feature", value_name="feature_value",
    )
    long_post["harmonization_state"] = "Post-ComBat"
    return pd.concat([long_pre, long_post], ignore_index=True)


if __name__ == "__main__":
    # Day-3 CLI entrypoint — runs with default paths against `data/raw/mri/`.
    # Expects `data/raw/mri/sites.csv` with columns `subject_id, site`.
    # Argument parsing (argparse / click) will land in a later task.
    #   python -m src.pipelines.mri_pipeline
    run_pipeline()