Blame for backend/utils.py

from pathlib import Path
import json
import ncempy.io as nio
import tifffile
import cv2
import numpy as np
import re
import h5py
import xml.etree.ElementTree as ET

def _convert_to_nm(value, unit):
    """Convert a pixel size value from given unit to nanometers."""
    unit = unit.lower().strip()
    conversions = {
        'pm': 1e-3, 'picometer': 1e-3, 'picometers': 1e-3,
        'a': 0.1, 'å': 0.1, 'angstrom': 0.1, 'angstroms': 0.1, 'ang': 0.1,
        'nm': 1, 'nanometer': 1, 'nanometers': 1,
        'um': 1e3, 'µm': 1e3, 'micrometer': 1e3, 'micrometers': 1e3, 'micron': 1e3, 'microns': 1e3,
        'mm': 1e6, 'millimeter': 1e6, 'millimeters': 1e6,
        'cm': 1e7, 'centimeter': 1e7, 'centimeters': 1e7,
        'm': 1e9, 'meter': 1e9, 'meters': 1e9,
        'inch': 2.54e7, 'inches': 2.54e7,
    }
    factor = conversions.get(unit)
    if factor is not None:
        return value * factor
    return None


def _decode_tiff_tag_texts(value):
    """Yield searchable text variants from TIFF tag values, including raw bytes."""
    if value is None:
        return []

    texts = []
    if isinstance(value, bytes):
        for encoding in ("latin1", "utf-8", "utf-16le"):
            try:
                decoded = value.decode(encoding, errors="ignore")
            except Exception:
                continue
            if decoded:
                texts.append(decoded)
    else:
        try:
            text = str(value)
        except Exception:
            text = ""
        if text:
            texts.append(text)

    variants = []
    seen = set()
    for text in texts:
        for candidate in (text, text.replace("\x00", "")):
            if candidate and candidate not in seen:
                seen.add(candidate)
                variants.append(candidate)
    return variants


def _normalize_h5_attr(value):
    if isinstance(value, bytes):
        return value.decode("utf-8", errors="ignore")
    return value


def _coerce_float(value):
    try:
        return float(value)
    except (TypeError, ValueError):
        return None


def _emd_dataset_to_2d(raw):
    """Reduce common EMD dataset layouts to a 2D image plane."""
    arr = np.asarray(raw)
    if arr.ndim < 2:
        return None

    arr = np.squeeze(arr)
    if arr.ndim < 2:
        return None

    while arr.ndim > 2:
        # Prefer collapsing the smallest axis first:
        # (y, x, 1) -> squeeze to (y, x)
        # (frames, y, x) -> take first frame
        # (y, x, channels) -> take first channel
        axis = int(np.argmin(arr.shape))
        arr = np.take(arr, indices=0, axis=axis)

    return arr


def _normalize_to_uint8(raw):
    arr = np.asarray(raw)
    if arr.size == 0:
        return None
    norm = cv2.normalize(arr, None, 0, 255, cv2.NORM_MINMAX)
    return norm.astype(np.uint8)


def _iter_emd_image_groups(h5_file):
    """Yield candidate EMD image groups in Velox and Berkeley-style layouts."""
    if 'Data' in h5_file and 'Image' in h5_file['Data']:
        for key, grp in h5_file['Data']['Image'].items():
            if isinstance(grp, h5py.Group) and 'Data' in grp and isinstance(grp['Data'], h5py.Dataset):
                yield "velox", key, grp, 'Data'

    if 'data' in h5_file:
        for key, grp in h5_file['data'].items():
            if isinstance(grp, h5py.Group) and 'data' in grp and isinstance(grp['data'], h5py.Dataset):
                yield "berkeley", key, grp, 'data'


def _read_emd_metadata_text(group):
    dataset = group.get('Metadata')
    if not isinstance(dataset, h5py.Dataset):
        return None

    try:
        raw = bytes(np.asarray(dataset).reshape(-1).tolist())
    except Exception:
        return None

    text = raw.decode("utf-8", errors="ignore").strip("\x00 \n\r\t")
    return text or None


def _extract_emd_pixel_size_from_metadata_text(metadata_text):
    if not metadata_text:
        return None

    try:
        payload = json.loads(metadata_text)
    except json.JSONDecodeError:
        payload = None

    if payload is not None:
        stack = [payload]
        while stack:
            node = stack.pop()
            if isinstance(node, dict):
                pixel_size = node.get("PixelSize")
                if isinstance(pixel_size, dict):
                    for axis_key, unit_key in (("width", "PixelUnitX"), ("height", "PixelUnitY"), ("x", "PixelUnitX"), ("y", "PixelUnitY")):
                        value = _coerce_float(pixel_size.get(axis_key))
                        unit = node.get(unit_key) or node.get("PixelUnit") or pixel_size.get("unit")
                        if value is not None and unit:
                            nm = _convert_to_nm(value, str(unit))
                            if nm and nm > 0:
                                return nm
                stack.extend(node.values())
            elif isinstance(node, list):
                stack.extend(node)

    match = re.search(
        r'"PixelSize"\s*:\s*\{[^{}]*"width"\s*:\s*"?(?P<value>[\d.eE+-]+)"?[^{}]*\}[^{}]*"PixelUnitX"\s*:\s*"(?P<unit>[^"]+)"',
        metadata_text,
        re.IGNORECASE | re.DOTALL,
    )
    if match:
        value = _coerce_float(match.group("value"))
        unit = match.group("unit")
        if value is not None:
            nm = _convert_to_nm(value, unit)
            if nm and nm > 0:
                return nm

    return None


def read_emd_pixel_size(image_path: Path):
    """Read pixel size from common EMD metadata layouts and Velox JSON metadata."""
    try:
        with h5py.File(str(image_path), 'r') as f:
            for fmt, key, grp, _ in _iter_emd_image_groups(f):
                if fmt == "velox":
                    pixel_size_group = grp.get('PixelSize')
                    if isinstance(pixel_size_group, h5py.Group):
                        width = _coerce_float(_normalize_h5_attr(pixel_size_group.attrs.get('width')))
                        unit = _normalize_h5_attr(pixel_size_group.attrs.get('unit', 'm'))
                        if width is not None:
                            nm = _convert_to_nm(width, str(unit))
                            if nm and nm > 0:
                                return nm

                    dimension_group = grp.get('Dimension')
                    if isinstance(dimension_group, h5py.Group):
                        for dim in dimension_group.values():
                            if not isinstance(dim, h5py.Dataset):
                                continue
                            scale = _coerce_float(_normalize_h5_attr(dim.attrs.get('Scale')))
                            unit = _normalize_h5_attr(dim.attrs.get('Unit', ''))
                            if scale is None or not unit:
                                continue
                            nm = _convert_to_nm(scale, str(unit))
                            if nm and nm > 0:
                                return nm

                    metadata_text = _read_emd_metadata_text(grp)
                    metadata_nm = _extract_emd_pixel_size_from_metadata_text(metadata_text)
                    if metadata_nm:
                        return metadata_nm

                if fmt == "berkeley":
                    dim1 = grp.get('dim1')
                    if isinstance(dim1, h5py.Dataset):
                        dim1_data = dim1[:]
                        if len(dim1_data) > 1:
                            delta = _coerce_float(dim1_data[1] - dim1_data[0])
                            if delta is not None and delta > 0:
                                return delta
    except Exception as e:
        print(f"Error reading EMD metadata: {e}")

    return None


def read_emd_image(image_path: Path):
    """Read and normalize the best available 2D image plane from an EMD file."""
    try:
        with h5py.File(str(image_path), 'r') as f:
            best_plane = None
            best_area = -1

            for _, _, grp, data_key in _iter_emd_image_groups(f):
                raw = grp[data_key][:]
                plane = _emd_dataset_to_2d(raw)
                if plane is None:
                    continue

                area = int(plane.shape[0]) * int(plane.shape[1])
                if area > best_area:
                    best_plane = plane
                    best_area = area

            if best_plane is not None:
                return _normalize_to_uint8(best_plane)
    except Exception as e:
        print(f"Error reading EMD image: {e}")

    return None


def _read_tiff_pixel_size(image_path):
    """
    Try all known methods to extract pixel size from a TIFF file.
    Returns (pixel_size_nm, method_string) or (None, None).
    """
    with tifffile.TiffFile(str(image_path)) as tif:

        # --- Method 1: ImageJ metadata ---
        meta = getattr(tif, 'imagej_metadata', None)
        if meta:
            # 1a. Direct 'spacing' + 'unit'
            if 'unit' in meta and 'spacing' in meta:
                spacing = float(meta['spacing'])
                nm = _convert_to_nm(spacing, meta['unit'])
                if nm is not None and nm > 0:
                    print(f"[TIFF scale] ImageJ spacing={spacing} {meta['unit']} → {nm} nm/px")
                    return nm, "metadata_imagej"

            # 1b. ImageJ 'info' string (often contains scale info as key=value pairs)
            info_str = meta.get('Info') or meta.get('info') or ''
            if info_str:
                # Try "Pixel Width = 0.245 nm" or "Scale = 0.5 um" patterns
                for pattern in [
                    r'pixel\s*(?:width|size)\s*[=:]\s*([\d.eE+-]+)\s*(nm|µm|um|mm|m)\b',
                    r'scale\s*[=:]\s*([\d.eE+-]+)\s*(nm|µm|um|mm|m)\b',
                ]:
                    m = re.search(pattern, info_str, re.IGNORECASE)
                    if m:
                        val = float(m.group(1))
                        nm = _convert_to_nm(val, m.group(2))
                        if nm and nm > 0:
                            print(f"[TIFF scale] ImageJ info field: {m.group(0)} → {nm} nm/px")
                            return nm, "metadata_imagej_info"

        # --- Method 2: OME-TIFF (XML in ImageDescription) ---
        if tif.pages:
            desc_tag = tif.pages[0].tags.get('ImageDescription')
            if desc_tag:
                desc = str(desc_tag.value)
                if desc.strip().startswith('<?xml') or '<OME' in desc:
                    try:
                        # Strip XML namespace for simpler parsing
                        desc_clean = re.sub(r'\sxmlns="[^"]+"', '', desc, count=1)
                        root = ET.fromstring(desc_clean)
                        # Look for <Pixels PhysicalSizeX="..." PhysicalSizeXUnit="...">
                        for elem in root.iter():
                            if 'Pixels' in elem.tag:
                                ps_x = elem.attrib.get('PhysicalSizeX')
                                ps_unit = elem.attrib.get('PhysicalSizeXUnit', 'µm')  # OME default is µm
                                if ps_x:
                                    val = float(ps_x)
                                    nm = _convert_to_nm(val, ps_unit)
                                    if nm and nm > 0:
                                        print(f"[TIFF scale] OME-TIFF PhysicalSizeX={ps_x} {ps_unit} → {nm} nm/px")
                                        return nm, "metadata_ome"
                    except ET.ParseError:
                        pass

        # --- Method 3: FEI / Thermo Fisher / DigitalMicrograph custom TIFF tags ---
        if tif.pages:
            page = tif.pages[0]
            for tag_id in (34682, 65027):
                fei_tag = page.tags.get(tag_id)
                if not fei_tag:
                    continue

                for fei_meta in _decode_tiff_tag_texts(fei_tag.value):
                    # Legacy FEI metadata stores PixelWidth in meters.
                    m = re.search(r'PixelWidth\s*=\s*([\d.eE+-]+)', fei_meta)
                    if m:
                        val_m = float(m.group(1))
                        nm = val_m * 1e9
                        if nm > 0:
                            print(f"[TIFF scale] FEI tag {tag_id} PixelWidth={val_m} m → {nm} nm/px")
                            return nm, f"metadata_fei_{tag_id}"

                    # DigitalMicrograph/JEOL headers often store PixelSizeX/Y as UTF-16-ish bytes.
                    for pattern in [
                        r'PixelSize(?:[XY])?[^0-9-]{0,80}(-?[\d.]+(?:e[-+]?\d+)?)\s*(nm|µm|um|mm|m)\b',
                        r'Pixel(?:Width|Height)(?:[XY])?[^0-9-]{0,80}(-?[\d.]+(?:e[-+]?\d+)?)\s*(nm|µm|um|mm|m)\b',
                    ]:
                        m = re.search(pattern, fei_meta, re.IGNORECASE | re.DOTALL)
                        if m:
                            val = float(m.group(1))
                            nm = _convert_to_nm(val, m.group(2))
                            if nm and nm > 0:
                                print(f"[TIFF scale] Private tag {tag_id}: {m.group(0)} → {nm} nm/px")
                                return nm, f"metadata_private_{tag_id}"

        # --- Method 4: ImageDescription free-text patterns ---
        if tif.pages:
            desc_tag = tif.pages[0].tags.get('ImageDescription')
            if desc_tag:
                desc = str(desc_tag.value)
                for pattern in [
                    r'pixel\s*size\s*[=:]\s*([\d.eE+-]+)\s*(nm|µm|um|mm|m)\b',
                    r'scale\s*[=:]\s*([\d.eE+-]+)\s*(nm|µm|um|mm|m)\b',
                    r'resolution\s*[=:]\s*([\d.eE+-]+)\s*(nm|µm|um|mm|m)\s*/\s*(?:pixel|px)',
                ]:
                    m = re.search(pattern, desc, re.IGNORECASE)
                    if m:
                        val = float(m.group(1))
                        nm = _convert_to_nm(val, m.group(2))
                        if nm and nm > 0:
                            print(f"[TIFF scale] ImageDescription: {m.group(0)} → {nm} nm/px")
                            return nm, "metadata_tiff_description"

        # --- Method 5: Standard TIFF resolution tags ---
        # CAUTION: most TIFFs have default 72 or 96 dpi, which is meaningless for microscopy.
        # Only use this if the result gives a plausible nm/pixel for TEM/SEM (0.01 ~ 100 nm/px).
        if tif.pages:
            page = tif.pages[0]
            tags = page.tags
            x_res_tag = tags.get('XResolution')
            res_unit_tag = tags.get('ResolutionUnit')
            if x_res_tag is not None:
                x_res_val = x_res_tag.value
                pixels_per_unit = None
                if isinstance(x_res_val, tuple) and len(x_res_val) == 2 and x_res_val[0] > 0:
                    pixels_per_unit = x_res_val[0] / x_res_val[1]
                elif isinstance(x_res_val, (int, float)) and x_res_val > 0:
                    pixels_per_unit = float(x_res_val)

                if pixels_per_unit and pixels_per_unit > 0:
                    res_unit = res_unit_tag.value if res_unit_tag else 1
                    nm_per_px = None
                    if res_unit == 3:  # centimeter
                        nm_per_px = 1e7 / pixels_per_unit
                    elif res_unit == 2:  # inch
                        nm_per_px = 2.54e7 / pixels_per_unit

                    # Only trust if result is in plausible microscopy range
                    if nm_per_px is not None and 0.01 <= nm_per_px <= 1000:
                        print(f"[TIFF scale] Resolution tag: {pixels_per_unit} px/unit (unit={res_unit}) → {nm_per_px} nm/px")
                        return nm_per_px, "metadata_tiff_resolution"
                    elif nm_per_px is not None:
                        print(f"[TIFF scale] Resolution tag gives {nm_per_px} nm/px — outside plausible range, ignoring (likely default DPI)")

        # --- Method 6: Check all TIFF tags for anything scale-related ---
        if tif.pages:
            for tag in tif.pages[0].tags.values():
                for tag_str in _decode_tiff_tag_texts(tag.value):
                    if len(tag_str) > 20000:
                        continue  # skip huge binary blobs
                    for pattern in [
                        r'(?:pixel|image)\s*(?:width|size|scale|resolution)\s*[=:]\s*([\d.eE+-]+)\s*(nm|µm|um|mm|m)\b',
                        r'(?:pixel|image)\s*(?:width|size|scale|resolution)[^0-9-]{0,80}(-?[\d.]+(?:e[-+]?\d+)?)\s*(nm|µm|um|mm|m)\b',
                    ]:
                        m = re.search(pattern, tag_str, re.IGNORECASE | re.DOTALL)
                        if m:
                            val = float(m.group(1))
                            nm = _convert_to_nm(val, m.group(2))
                            if nm and nm > 0:
                                print(f"[TIFF scale] Tag {tag.name}: {m.group(0)} → {nm} nm/px")
                                return nm, f"metadata_tiff_tag_{tag.name}"

    return None, None


# --- Burned-in scale detection (OCR) ----------------------------------------
# Many camera exports (e.g. Velox/Ceta JPG/PNG) carry no embedded calibration
# but print it into the image: a metadata footer ("Pixel size 156.1 pm",
# "Fov 639.3 nm") and/or a drawn scale bar with a "200 nm" label. We read those
# with Tesseract OCR. Everything degrades silently when Tesseract/pytesseract is
# not installed, so the app keeps working (falling back to manual calibration).

_PLAUSIBLE_NM_PER_PX = (0.001, 1000.0)  # sane TEM/SEM range


def _tesseract_ready():
    """Return True if pytesseract + the Tesseract binary are usable; point
    pytesseract at common install locations (incl. Windows) when needed."""
    import os
    import shutil
    try:
        import pytesseract
    except Exception:
        return False

    if shutil.which("tesseract"):
        return True

    for candidate in (
        r"C:\Program Files\Tesseract-OCR\tesseract.exe",
        r"C:\Program Files (x86)\Tesseract-OCR\tesseract.exe",
        "/opt/homebrew/bin/tesseract",
        "/usr/local/bin/tesseract",
        "/usr/bin/tesseract",
    ):
        if os.path.exists(candidate):
            pytesseract.pytesseract.tesseract_cmd = candidate
            return True
    return False


def _ocr_text(gray, psm=6):
    """OCR a grayscale crop, upscaling small crops so digits are legible."""
    import pytesseract
    if gray is None or gray.size == 0:
        return ""
    h, w = gray.shape[:2]
    if max(h, w) < 2500:
        gray = cv2.resize(gray, (w * 2, h * 2), interpolation=cv2.INTER_CUBIC)
    try:
        return pytesseract.image_to_string(gray, config=f"--psm {psm}")
    except Exception as e:
        print(f"[Scale] OCR error: {e}")
        return ""


def _plausible(nm_per_px):
    lo, hi = _PLAUSIBLE_NM_PER_PX
    return nm_per_px is not None and lo <= nm_per_px <= hi


def _parse_footer_calibration(text, image_width):
    """
    Parse a metadata footer into nm/pixel. Two independent estimates:
      A = an explicit pixel-size token in pm  (e.g. "156.1 pm")
      B = field-of-view length / image width  (e.g. "639.3 nm" / 4096)
    Prefer A when present; use B otherwise. When both exist they cross-check
    each other, which guards against OCR misreads.
    """
    norm = " ".join(text.split())
    tokens = re.findall(r'([\d]+\.?[\d]*)\s*(pm|nm|µm|um|microns?)\b', norm, re.IGNORECASE)
    pm_vals, len_nm_vals = [], []
    for value, unit in tokens:
        u = unit.lower()
        nm = _convert_to_nm(float(value), u)
        if nm is None:
            continue
        if u == 'pm':
            pm_vals.append(nm)
        else:
            len_nm_vals.append(nm)

    a = pm_vals[0] if pm_vals else None              # explicit pixel size (pm)
    # Field of view is the largest spatial length in the footer.
    b = (max(len_nm_vals) / image_width) if (len_nm_vals and image_width) else None

    if _plausible(a):
        if b is not None and abs(a - b) / max(a, b) > 0.35:
            # A and B disagree strongly — A is the labelled pixel size, keep it
            # but note the mismatch for transparency.
            return a, "ocr_pixel_size", f"OCR footer pixel size ({a:.4f} nm/px); FOV cross-check differs"
        return a, "ocr_pixel_size", f"OCR footer pixel size: {a:.4f} nm/px"
    if _plausible(b):
        return b, "ocr_fov", f"OCR footer field-of-view / {image_width}px: {b:.4f} nm/px"
    return None, None, None


def _detect_scale_bar_line(img):
    """
    Tertiary fallback: find a drawn scale bar (a wide, thin bright/dark bar near
    the bottom) and OCR its length label. Strict on purpose — a wrong scale is
    worse than none — so it only returns a value when a unit label is actually
    read next to a plausibly-sized bar.
    """
    h, w = img.shape[:2]
    strip_top = int(h * 0.8)
    bottom = img[strip_top:, :]

    for invert in (False, True):
        b = (255 - bottom) if invert else bottom
        _, thresh = cv2.threshold(b, 200, 255, cv2.THRESH_BINARY)
        contours, _ = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

        best = None
        best_w = 0
        for cnt in contours:
            x, y, cw, ch = cv2.boundingRect(cnt)
            # Wide, thin, and not absurdly long (reject rod rows / artefacts).
            if cw > max(40, w * 0.03) and ch < 25 and cw / max(ch, 1) > 6 and cw < w * 0.45:
                if cw > best_w:
                    best_w = cw
                    best = (x, y + strip_top, cw, ch)
        if not best:
            continue

        lx, ly, lw, lh = best
        ty1, ty2 = max(0, ly - 90), min(h, ly + lh + 25)
        tx1, tx2 = max(0, lx - 40), min(w, lx + lw + 140)
        label_crop = img[ty1:ty2, tx1:tx2]
        for psm in (7, 11):
            label = _ocr_text(label_crop, psm=psm)
            m = re.search(r'([\d]+\.?[\d]*)\s*(pm|nm|µm|um|microns?)\b',
                          " ".join(label.split()), re.IGNORECASE)
            if not m:
                continue
            nm = _convert_to_nm(float(m.group(1)), m.group(2))
            if nm and lw > 0:
                ps = nm / lw
                if _plausible(ps):
                    return ps, {
                        "method": "ocr_scale_bar",
                        "scale_bar_coords": (lx, ly + lh // 2, lx + lw, ly + lh // 2),
                        "scale_bar_length_nm": nm,
                        "scale_bar_length_px": lw,
                        "description": f"OCR scale bar: {m.group(1)} {m.group(2)} over {lw}px",
                    }
    return None, {}


def detect_burned_in_scale(image_path: Path):
    """
    Auto-calibrate from burned-in text/scale bar in a raster image (JPG/PNG/TIF).
    Returns (pixel_size_nm, calibration_info) or (None, {}). Safe no-op without
    Tesseract.
    """
    if not _tesseract_ready():
        return None, {}

    try:
        img = cv2.imread(str(image_path), cv2.IMREAD_GRAYSCALE)
        if img is None:
            return None, {}
        h, w = img.shape[:2]

        # 1 & 2. Metadata footer text (pixel size / field of view).
        try:
            from autodetect_utils import _detect_footer_top
            footer_top = _detect_footer_top(img)
        except Exception:
            footer_top = h
        # OCR the detected footer band; if footer detection found nothing, fall
        # back to a single generous bottom slice. One OCR pass either way.
        footer_region = img[footer_top:, :] if footer_top < h else img[int(h * 0.88):, :]
        ps, method, desc = _parse_footer_calibration(_ocr_text(footer_region, psm=6), w)
        if ps is not None:
            print(f"[Scale] {desc}")
            return ps, {"method": method, "pixel_size_nm": ps, "description": desc,
                        "scale_bar_coords": None, "scale_bar_length_nm": None,
                        "scale_bar_length_px": None}

        # 3. Drawn scale bar + label.
        ps, info = _detect_scale_bar_line(img)
        if ps is not None:
            print(f"[Scale] {info.get('description')}")
            return ps, info
    except Exception as e:
        print(f"[Scale] burned-in detection failed: {e}")

    return None, {}


def get_pixel_size(image_path: Path):
    """
    Extract pixel size (nm/pixel) from image metadata or scale bar.
    Returns a tuple: (pixel_size_nm, calibration_info)
    calibration_info is a dict with details for visualization (e.g., scale bar coordinates).
    """
    ext = image_path.suffix.lower()
    pixel_size = None
    method = "unknown"
    scale_bar_coords = None  # (x1, y1, x2, y2)
    scale_bar_length_nm = None
    best_line_width_px = None

    # 1. Try Metadata
    try:
        if ext in ['.dm3', '.dm4']:
            dm = nio.read(str(image_path))
            if 'pixelSize' in dm:
                pixel_size = dm['pixelSize'][0]
                method = "metadata_dm"
                print(f"[Scale] DM metadata: {pixel_size} nm/px")

        elif ext == '.emd':
            pixel_size = read_emd_pixel_size(image_path)
            if pixel_size is not None:
                method = "metadata_emd"

        elif ext in ['.tif', '.tiff']:
            pixel_size, method = _read_tiff_pixel_size(image_path)
            if pixel_size is None:
                method = "unknown"

    except Exception as e:
        print(f"Error extracting metadata: {e}")
        import traceback
        traceback.print_exc()

    # 2. Fallback: read calibration burned into the image (footer text / scale
    # bar) via OCR. Only when metadata gave nothing, so calibrated formats are
    # not slowed down.
    description = None
    if pixel_size is None:
        ocr_ps, ocr_info = detect_burned_in_scale(image_path)
        if ocr_ps is not None:
            pixel_size = ocr_ps
            method = ocr_info.get("method", "ocr")
            scale_bar_coords = ocr_info.get("scale_bar_coords")
            scale_bar_length_nm = ocr_info.get("scale_bar_length_nm")
            best_line_width_px = ocr_info.get("scale_bar_length_px")
            description = ocr_info.get("description")

    if pixel_size is None:
        method = "uncalibrated"
        print(f"[Scale] No calibration found for {image_path.name}")
    else:
        print(f"[Scale] Final: {pixel_size} nm/px via {method}")

    calibration_info = {
        "method": method,
        "scale_bar_coords": scale_bar_coords,
        "scale_bar_length_nm": scale_bar_length_nm,
        "scale_bar_length_px": best_line_width_px if scale_bar_coords else None
    }
    if description:
        calibration_info["description"] = description
    if pixel_size is None:
        calibration_info["warning"] = "No calibration found. Measurements are using a placeholder scale until you calibrate manually."

    return pixel_size, calibration_info