phasorpy 0.7__cp314-cp314-win_amd64.whl → 0.8__cp314-cp314-win_amd64.whl

phasorpy/filter.py

@@ -0,0 +1,966 @@
+"""Filter signals and phasor coordinates.
+
+The ``phasorpy.filter`` module provides functions to filter
+
+- phasor coordinates:
+
+  - :py:func:`phasor_filter_gaussian` (not implemented yet)
+  - :py:func:`phasor_filter_median`
+  - :py:func:`phasor_filter_pawflim`
+  - :py:func:`phasor_threshold`
+
+- signals:
+
+  - :py:func:`signal_filter_ncpca`
+    (noise-corrected principal component analysis)
+  - :py:func:`signal_filter_svd` (spectral vector denoise)
+  - :py:func:`signal_filter_median` (not implemented yet)
+
+"""
+
+from __future__ import annotations
+
+__all__ = [
+    # 'signal_filter_gaussian',
+    'phasor_filter_median',
+    'phasor_filter_pawflim',
+    'phasor_threshold',
+    # 'signal_filter_median',
+    'signal_filter_ncpca',
+    'signal_filter_svd',
+]
+
+import math
+from collections.abc import Sequence
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from ._typing import Any, NDArray, ArrayLike, DTypeLike, Literal
+
+import numpy
+
+from ._phasorpy import (
+    _median_filter_2d,
+    _phasor_from_signal_vector,
+    _phasor_threshold_closed,
+    _phasor_threshold_mean_closed,
+    _phasor_threshold_mean_open,
+    _phasor_threshold_nan,
+    _phasor_threshold_open,
+    _signal_denoise_vector,
+)
+from ._utils import parse_harmonic, parse_skip_axis
+from .utils import number_threads
+
+
+def phasor_filter_median(
+    mean: ArrayLike,
+    real: ArrayLike,
+    imag: ArrayLike,
+    /,
+    *,
+    repeat: int = 1,
+    size: int = 3,
+    skip_axis: int | Sequence[int] | None = None,
+    use_scipy: bool = False,
+    num_threads: int | None = None,
+    **kwargs: Any,
+) -> tuple[NDArray[Any], NDArray[Any], NDArray[Any]]:
+    """Return median-filtered phasor coordinates.
+
+    By default, apply a NaN-aware median filter independently to the real
+    and imaginary components of phasor coordinates once with a kernel size of 3
+    multiplied by the number of dimensions of the input arrays. Return the
+    intensity unchanged.
+
+    Parameters
+    ----------
+    mean : array_like
+        Intensity of phasor coordinates.
+    real : array_like
+        Real component of phasor coordinates to be filtered.
+    imag : array_like
+        Imaginary component of phasor coordinates to be filtered.
+    repeat : int, optional
+        Number of times to apply median filter. The default is 1.
+    size : int, optional
+        Size of median filter kernel. The default is 3.
+    skip_axis : int or sequence of int, optional
+        Axes in `mean` to exclude from filter.
+        By default, all axes except harmonics are included.
+    use_scipy : bool, optional
+        Use :py:func:`scipy.ndimage.median_filter`.
+        This function has undefined behavior if the input arrays contain
+        NaN values but is faster when filtering more than 2 dimensions.
+        See `issue #87 <https://github.com/phasorpy/phasorpy/issues/87>`_.
+    num_threads : int, optional
+        Number of OpenMP threads to use for parallelization.
+        Applies to filtering in two dimensions when not using scipy.
+        By default, multi-threading is disabled.
+        If zero, up to half of logical CPUs are used.
+        OpenMP may not be available on all platforms.
+    **kwargs
+        Optional arguments passed to :py:func:`scipy.ndimage.median_filter`.
+
+    Returns
+    -------
+    mean : ndarray
+        Unchanged intensity of phasor coordinates.
+    real : ndarray
+        Filtered real component of phasor coordinates.
+    imag : ndarray
+        Filtered imaginary component of phasor coordinates.
+
+    Raises
+    ------
+    ValueError
+        If `repeat` is less than 0.
+        If `size` is less than 1.
+        The array shapes of `mean`, `real`, and `imag` do not match.
+
+    Examples
+    --------
+    Apply three times a median filter with a kernel size of three:
+
+    >>> mean, real, imag = phasor_filter_median(
+    ...     [[1, 2, 3], [4, 5, 6], [7, 8, 9]],
+    ...     [[0.0, 0.0, 0.0], [0.5, 0.5, 0.5], [0.2, 0.2, 0.2]],
+    ...     [[0.3, 0.3, 0.3], [0.6, math.nan, 0.6], [0.4, 0.4, 0.4]],
+    ...     size=3,
+    ...     repeat=3,
+    ... )
+    >>> mean
+    array([[1, 2, 3],
+           [4, 5, 6],
+           [7, 8, 9]])
+    >>> real
+    array([[0, 0, 0],
+           [0.2, 0.2, 0.2],
+           [0.2, 0.2, 0.2]])
+    >>> imag
+    array([[0.3, 0.3, 0.3],
+           [0.4, nan, 0.4],
+           [0.4, 0.4, 0.4]])
+
+    """
+    if repeat < 0:
+        raise ValueError(f'{repeat=} < 0')
+    if size < 1:
+        raise ValueError(f'{size=} < 1')
+    if size == 1:
+        # no need to filter
+        repeat = 0
+
+    mean = numpy.asarray(mean)
+    if use_scipy or repeat == 0:  # or using nD numpy filter
+        real = numpy.asarray(real)
+    elif isinstance(real, numpy.ndarray) and real.dtype == numpy.float32:
+        real = real.copy()
+    else:
+        real = numpy.asarray(real, dtype=numpy.float64, copy=True)
+    if use_scipy or repeat == 0:  # or using nD numpy filter
+        imag = numpy.asarray(imag)
+    elif isinstance(imag, numpy.ndarray) and imag.dtype == numpy.float32:
+        imag = imag.copy()
+    else:
+        imag = numpy.asarray(imag, dtype=numpy.float64, copy=True)
+
+    if mean.shape != real.shape[-mean.ndim if mean.ndim else 1 :]:
+        raise ValueError(f'{mean.shape=} != {real.shape=}')
+    if real.shape != imag.shape:
+        raise ValueError(f'{real.shape=} != {imag.shape=}')
+
+    prepend_axis = mean.ndim + 1 == real.ndim
+    _, axes = parse_skip_axis(skip_axis, mean.ndim, prepend_axis)
+
+    # in case mean is also filtered
+    # if prepend_axis:
+    #     mean = numpy.expand_dims(mean, axis=0)
+    # ...
+    # if prepend_axis:
+    #     mean = numpy.asarray(mean[0])
+
+    if repeat == 0:
+        # no need to call filter
+        return mean, real, imag
+
+    if use_scipy:
+        # use scipy NaN-unaware fallback
+        from scipy.ndimage import median_filter
+
+        kwargs.pop('axes', None)
+
+        for _ in range(repeat):
+            real = median_filter(real, size=size, axes=axes, **kwargs)
+            imag = median_filter(imag, size=size, axes=axes, **kwargs)
+
+        return mean, numpy.asarray(real), numpy.asarray(imag)
+
+    if len(axes) != 2:
+        # n-dimensional median filter using numpy
+        from numpy.lib.stride_tricks import sliding_window_view
+
+        kernel_shape = tuple(
+            size if i in axes else 1 for i in range(real.ndim)
+        )
+        pad_width = [
+            (s // 2, s // 2) if s > 1 else (0, 0) for s in kernel_shape
+        ]
+        axis = tuple(range(-real.ndim, 0))
+
+        nan_mask = numpy.isnan(real)
+        for _ in range(repeat):
+            real = numpy.pad(real, pad_width, mode='edge')
+            real = sliding_window_view(real, kernel_shape)
+            real = numpy.nanmedian(real, axis=axis)
+            real = numpy.where(nan_mask, numpy.nan, real)
+
+        nan_mask = numpy.isnan(imag)
+        for _ in range(repeat):
+            imag = numpy.pad(imag, pad_width, mode='edge')
+            imag = sliding_window_view(imag, kernel_shape)
+            imag = numpy.nanmedian(imag, axis=axis)
+            imag = numpy.where(nan_mask, numpy.nan, imag)
+
+        return mean, real, imag
+
+    # 2-dimensional median filter using optimized Cython implementation
+    num_threads = number_threads(num_threads)
+
+    buffer = numpy.empty(
+        tuple(real.shape[axis] for axis in axes), dtype=real.dtype
+    )
+
+    for index in numpy.ndindex(
+        *[real.shape[ax] for ax in range(real.ndim) if ax not in axes]
+    ):
+        index_list: list[int | slice] = list(index)
+        for ax in axes:
+            index_list = index_list[:ax] + [slice(None)] + index_list[ax:]
+        full_index = tuple(index_list)
+
+        _median_filter_2d(real[full_index], buffer, size, repeat, num_threads)
+        _median_filter_2d(imag[full_index], buffer, size, repeat, num_threads)
+
+    return mean, real, imag
+
+
+def phasor_filter_pawflim(
+    mean: ArrayLike,
+    real: ArrayLike,
+    imag: ArrayLike,
+    /,
+    *,
+    sigma: float = 2.0,
+    levels: int = 1,
+    harmonic: Sequence[int] | None = None,
+    skip_axis: int | Sequence[int] | None = None,
+) -> tuple[NDArray[Any], NDArray[Any], NDArray[Any]]:
+    """Return pawFLIM wavelet-filtered phasor coordinates.
+
+    This function must only be used with calibrated, unprocessed phasor
+    coordinates obtained from FLIM data. The coordinates must not be filtered,
+    thresholded, or otherwise pre-processed.
+
+    The pawFLIM wavelet filter is described in [1]_.
+
+    Parameters
+    ----------
+    mean : array_like
+        Intensity of phasor coordinates.
+    real : array_like
+        Real component of phasor coordinates to be filtered.
+        Must have at least two harmonics in the first axis.
+    imag : array_like
+        Imaginary component of phasor coordinates to be filtered.
+        Must have at least two harmonics in the first axis.
+    sigma : float, optional
+        Significance level to test difference between two phasors.
+        Given in terms of the equivalent 1D standard deviations.
+        sigma=2 corresponds to ~95% (or 5%) significance.
+    levels : int, optional
+        Number of levels for wavelet decomposition.
+        Controls the maximum averaging area, which has a length of
+        :math:`2^level`.
+    harmonic : sequence of int or None, optional
+        Harmonics included in first axis of `real` and `imag`.
+        If None (default), the first axis of `real` and `imag` contains lower
+        harmonics starting at and increasing by one.
+        All harmonics must have a corresponding half or double harmonic.
+    skip_axis : int or sequence of int, optional
+        Axes in `mean` to exclude from filter.
+        By default, all axes except harmonics are included.
+
+    Returns
+    -------
+    mean : ndarray
+        Unchanged intensity of phasor coordinates.
+    real : ndarray
+        Filtered real component of phasor coordinates.
+    imag : ndarray
+        Filtered imaginary component of phasor coordinates.
+
+    Raises
+    ------
+    ValueError
+        If `level` is less than 0.
+        The array shapes of `mean`, `real`, and `imag` do not match.
+        If `real` and `imag` have no harmonic axis.
+        Number of harmonics in `harmonic` is less than 2 or does not match
+        the first axis of `real` and `imag`.
+        Not all harmonics in `harmonic` have a corresponding half
+        or double harmonic.
+
+    References
+    ----------
+    .. [1] Silberberg M, and Grecco H. `pawFLIM: reducing bias and
+      uncertainty to enable lower photon count in FLIM experiments
+      <https://doi.org/10.1088/2050-6120/aa72ab>`_.
+      *Methods Appl Fluoresc*, 5(2): 024016 (2017)
+
+    Examples
+    --------
+    Apply a pawFLIM wavelet filter with four significance levels (sigma)
+    and three decomposition levels:
+
+    >>> mean, real, imag = phasor_filter_pawflim(
+    ...     [[1, 1], [1, 1]],
+    ...     [[[0.5, 0.8], [0.5, 0.8]], [[0.2, 0.4], [0.2, 0.4]]],
+    ...     [[[0.5, 0.4], [0.5, 0.4]], [[0.4, 0.5], [0.4, 0.5]]],
+    ...     sigma=4,
+    ...     levels=3,
+    ...     harmonic=[1, 2],
+    ... )
+    >>> mean
+    array([[1, 1],
+           [1, 1]])
+    >>> real
+    array([[[0.65, 0.65],
+            [0.65, 0.65]],
+           [[0.3, 0.3],
+            [0.3, 0.3]]])
+    >>> imag
+    array([[[0.45, 0.45],
+            [0.45, 0.45]],
+           [[0.45, 0.45],
+            [0.45, 0.45]]])
+
+    """
+    from pawflim import pawflim  # type: ignore[import-untyped]
+
+    if levels < 0:
+        raise ValueError(f'{levels=} < 0')
+    if levels == 0:
+        return numpy.asarray(mean), numpy.asarray(real), numpy.asarray(imag)
+
+    mean = numpy.asarray(mean, dtype=numpy.float64, copy=True)
+    real = numpy.asarray(real, dtype=numpy.float64, copy=True)
+    imag = numpy.asarray(imag, dtype=numpy.float64, copy=True)
+
+    if mean.shape != real.shape[-mean.ndim if mean.ndim else 1 :]:
+        raise ValueError(f'{mean.shape=} != {real.shape=}')
+    if real.shape != imag.shape:
+        raise ValueError(f'{real.shape=} != {imag.shape=}')
+
+    has_harmonic_axis = mean.ndim + 1 == real.ndim
+    if not has_harmonic_axis:
+        raise ValueError('no harmonic axis')
+    if harmonic is None:
+        harmonics, _ = parse_harmonic('all', real.shape[0])
+    else:
+        harmonics, _ = parse_harmonic(harmonic, None)
+    if len(harmonics) < 2:
+        raise ValueError(
+            'at least two harmonics required, ' f'got {len(harmonics)}'
+        )
+    if len(harmonics) != real.shape[0]:
+        raise ValueError(
+            'number of harmonics does not match first axis of real and imag'
+        )
+
+    mean = numpy.asarray(numpy.nan_to_num(mean, copy=False))
+    real = numpy.asarray(numpy.nan_to_num(real, copy=False))
+    imag = numpy.asarray(numpy.nan_to_num(imag, copy=False))
+    real *= mean
+    imag *= mean
+
+    mean_expanded = numpy.broadcast_to(mean, real.shape).copy()
+    original_mean_expanded = mean_expanded.copy()
+    real_filtered = real.copy()
+    imag_filtered = imag.copy()
+
+    _, axes = parse_skip_axis(skip_axis, mean.ndim, True)
+
+    for index in numpy.ndindex(
+        *(
+            real.shape[ax]
+            for ax in range(real.ndim)
+            if ax not in axes and ax != 0
+        )
+    ):
+        index_list: list[int | slice] = list(index)
+        for ax in axes:
+            index_list = index_list[:ax] + [slice(None)] + index_list[ax:]
+        full_index = tuple(index_list)
+
+        processed_harmonics = set()
+
+        for h in harmonics:
+            if h in processed_harmonics and (
+                h * 4 in harmonics or h * 2 not in harmonics
+            ):
+                continue
+            if h * 2 not in harmonics:
+                raise ValueError(
+                    f'harmonic {h} does not have a corresponding half '
+                    f'or double harmonic in {harmonics}'
+                )
+            n = harmonics.index(h)
+            n2 = harmonics.index(h * 2)
+
+            complex_phasor = numpy.empty(
+                (3, *original_mean_expanded[n][full_index].shape),
+                dtype=numpy.complex128,
+            )
+            complex_phasor[0] = original_mean_expanded[n][full_index]
+            complex_phasor[1] = real[n][full_index] + 1j * imag[n][full_index]
+            complex_phasor[2] = (
+                real[n2][full_index] + 1j * imag[n2][full_index]
+            )
+
+            complex_phasor = pawflim(
+                complex_phasor, n_sigmas=sigma, levels=levels
+            )
+
+            for i, idx in enumerate([n, n2]):
+                if harmonics[idx] in processed_harmonics:
+                    continue
+                mean_expanded[idx][full_index] = complex_phasor[0].real
+                real_filtered[idx][full_index] = complex_phasor[i + 1].real
+                imag_filtered[idx][full_index] = complex_phasor[i + 1].imag
+
+            processed_harmonics.add(h)
+            processed_harmonics.add(h * 2)
+
+    with numpy.errstate(divide='ignore', invalid='ignore'):
+        real = numpy.asarray(numpy.divide(real_filtered, mean_expanded))
+        imag = numpy.asarray(numpy.divide(imag_filtered, mean_expanded))
+
+    return mean, real, imag
+
+
+def phasor_threshold(
+    mean: ArrayLike,
+    real: ArrayLike,
+    imag: ArrayLike,
+    /,
+    mean_min: ArrayLike | None = None,
+    mean_max: ArrayLike | None = None,
+    *,
+    real_min: ArrayLike | None = None,
+    real_max: ArrayLike | None = None,
+    imag_min: ArrayLike | None = None,
+    imag_max: ArrayLike | None = None,
+    phase_min: ArrayLike | None = None,
+    phase_max: ArrayLike | None = None,
+    modulation_min: ArrayLike | None = None,
+    modulation_max: ArrayLike | None = None,
+    open_interval: bool = False,
+    detect_harmonics: bool = True,
+    **kwargs: Any,
+) -> tuple[NDArray[Any], NDArray[Any], NDArray[Any]]:
+    """Return phasor coordinates with values outside interval replaced by NaN.
+
+    Interval thresholds can be set for mean intensity, real and imaginary
+    coordinates, and phase and modulation.
+    Phasor coordinates smaller than minimum thresholds or larger than maximum
+    thresholds are replaced with NaN.
+    No threshold is applied by default.
+    NaNs in `mean` or any `real` and `imag` harmonic are propagated to
+    `mean` and all harmonics in `real` and `imag`.
+
+    Parameters
+    ----------
+    mean : array_like
+        Intensity of phasor coordinates.
+    real : array_like
+        Real component of phasor coordinates.
+    imag : array_like
+        Imaginary component of phasor coordinates.
+    mean_min : array_like, optional
+        Lower threshold for mean intensity.
+    mean_max : array_like, optional
+        Upper threshold for mean intensity.
+    real_min : array_like, optional
+        Lower threshold for real coordinates.
+    real_max : array_like, optional
+        Upper threshold for real coordinates.
+    imag_min : array_like, optional
+        Lower threshold for imaginary coordinates.
+    imag_max : array_like, optional
+        Upper threshold for imaginary coordinates.
+    phase_min : array_like, optional
+        Lower threshold for phase angle.
+    phase_max : array_like, optional
+        Upper threshold for phase angle.
+    modulation_min : array_like, optional
+        Lower threshold for modulation.
+    modulation_max : array_like, optional
+        Upper threshold for modulation.
+    open_interval : bool, optional
+        If true, the interval is open, and the threshold values are
+        not included in the interval.
+        If false (default), the interval is closed, and the threshold values
+        are included in the interval.
+    detect_harmonics : bool, optional
+        By default, detect presence of multiple harmonics from array shapes.
+        If false, no harmonics are assumed to be present, and the function
+        behaves like a numpy universal function.
+    **kwargs
+        Optional `arguments passed to numpy universal functions
+        <https://numpy.org/doc/stable/reference/ufuncs.html#ufuncs-kwargs>`_.
+
+    Returns
+    -------
+    mean : ndarray
+        Thresholded intensity of phasor coordinates.
+    real : ndarray
+        Thresholded real component of phasor coordinates.
+    imag : ndarray
+        Thresholded imaginary component of phasor coordinates.
+
+    Examples
+    --------
+    Set phasor coordinates to NaN if mean intensity is smaller than 1.1:
+
+    >>> phasor_threshold([1, 2, 3], [0.1, 0.2, 0.3], [0.4, 0.5, 0.6], 1.1)
+    (array([nan, 2, 3]), array([nan, 0.2, 0.3]), array([nan, 0.5, 0.6]))
+
+    Set phasor coordinates to NaN if real component is smaller than 0.15 or
+    larger than 0.25:
+
+    >>> phasor_threshold(
+    ...     [1.0, 2.0, 3.0],
+    ...     [0.1, 0.2, 0.3],
+    ...     [0.4, 0.5, 0.6],
+    ...     real_min=0.15,
+    ...     real_max=0.25,
+    ... )
+    (array([nan, 2, nan]), array([nan, 0.2, nan]), array([nan, 0.5, nan]))
+
+    Apply NaNs to other input arrays:
+
+    >>> phasor_threshold(
+    ...     [numpy.nan, 2, 3], [0.1, 0.2, 0.3], [0.4, 0.5, numpy.nan]
+    ... )
+    (array([nan, 2, nan]), array([nan, 0.2, nan]), array([nan, 0.5, nan]))
+
+    """
+    threshold_mean_only = None
+    if mean_min is None:
+        mean_min = numpy.nan
+    else:
+        threshold_mean_only = True
+    if mean_max is None:
+        mean_max = numpy.nan
+    else:
+        threshold_mean_only = True
+    if real_min is None:
+        real_min = numpy.nan
+    else:
+        threshold_mean_only = False
+    if real_max is None:
+        real_max = numpy.nan
+    else:
+        threshold_mean_only = False
+    if imag_min is None:
+        imag_min = numpy.nan
+    else:
+        threshold_mean_only = False
+    if imag_max is None:
+        imag_max = numpy.nan
+    else:
+        threshold_mean_only = False
+    if phase_min is None:
+        phase_min = numpy.nan
+    else:
+        threshold_mean_only = False
+    if phase_max is None:
+        phase_max = numpy.nan
+    else:
+        threshold_mean_only = False
+    if modulation_min is None:
+        modulation_min = numpy.nan
+    else:
+        threshold_mean_only = False
+    if modulation_max is None:
+        modulation_max = numpy.nan
+    else:
+        threshold_mean_only = False
+
+    if detect_harmonics:
+        mean = numpy.asarray(mean)
+        real = numpy.asarray(real)
+        imag = numpy.asarray(imag)
+
+        shape = numpy.broadcast_shapes(mean.shape, real.shape, imag.shape)
+        ndim = len(shape)
+
+        has_harmonic_axis = (
+            # detect multi-harmonic in axis 0
+            mean.ndim + 1 == ndim
+            and real.shape == shape
+            and imag.shape == shape
+            and mean.shape == shape[-mean.ndim if mean.ndim else 1 :]
+        )
+    else:
+        has_harmonic_axis = False
+
+    if threshold_mean_only is None:
+        mean, real, imag = _phasor_threshold_nan(mean, real, imag, **kwargs)
+
+    elif threshold_mean_only:
+        mean_func = (
+            _phasor_threshold_mean_open
+            if open_interval
+            else _phasor_threshold_mean_closed
+        )
+        mean, real, imag = mean_func(
+            mean, real, imag, mean_min, mean_max, **kwargs
+        )
+
+    else:
+        func = (
+            _phasor_threshold_open
+            if open_interval
+            else _phasor_threshold_closed
+        )
+        mean, real, imag = func(
+            mean,
+            real,
+            imag,
+            mean_min,
+            mean_max,
+            real_min,
+            real_max,
+            imag_min,
+            imag_max,
+            phase_min,
+            phase_max,
+            modulation_min,
+            modulation_max,
+            **kwargs,
+        )
+
+    mean = numpy.asarray(mean)
+    real = numpy.asarray(real)
+    imag = numpy.asarray(imag)
+    if has_harmonic_axis and mean.ndim > 0:
+        # propagate NaN to all dimensions
+        mean = numpy.mean(mean, axis=0, keepdims=True)
+        mask = numpy.where(numpy.isnan(mean), numpy.nan, 1.0)
+        numpy.multiply(real, mask, out=real)
+        numpy.multiply(imag, mask, out=imag)
+        # remove harmonic dimension created by broadcasting
+        mean = numpy.asarray(numpy.asarray(mean)[0])
+
+    return mean, real, imag
+
+
+def signal_filter_svd(
+    signal: ArrayLike,
+    /,
+    spectral_vector: ArrayLike | None = None,
+    *,
+    axis: int = -1,
+    harmonic: int | Sequence[int] | Literal['all'] | str | None = None,
+    sigma: float = 0.05,
+    vmin: float | None = None,
+    dtype: DTypeLike | None = None,
+    num_threads: int | None = None,
+) -> NDArray[Any]:
+    """Return spectral-vector-denoised signal.
+
+    The spectral vector denoising algorithm is based on a Gaussian weighted
+    average calculation, with weights obtained in n-dimensional Chebyshev or
+    Fourier space [2]_.
+
+    Parameters
+    ----------
+    signal : array_like
+        Hyperspectral data to be denoised.
+        A minimum of three samples are required along `axis`.
+        The samples must be uniformly spaced.
+    spectral_vector : array_like, optional
+        Spectral vector.
+        For example, phasor coordinates, PCA projected phasor coordinates,
+        or Chebyshev coefficients.
+        Must be of the same shape as `signal` with `axis` removed and an axis
+        containing spectral space appended.
+        If None (default), phasor coordinates are calculated at specified
+        `harmonic`.
+    axis : int, optional, default: -1
+        Axis over which `spectral_vector` is computed if not provided.
+        The default is the last axis (-1).
+    harmonic : int, sequence of int, or 'all', optional
+        Harmonics to include in calculating `spectral_vector`.
+        If `'all'`, include all harmonics for `signal` samples along `axis`.
+        Else, harmonics must be at least one and no larger than half the
+        number of `signal` samples along `axis`.
+        The default is the first harmonic (fundamental frequency).
+        A minimum of `harmonic * 2 + 1` samples are required along `axis`
+        to calculate correct phasor coordinates at `harmonic`.
+    sigma : float, default: 0.05
+        Width of Gaussian filter in spectral vector space.
+        Weighted averages are calculated using the spectra of signal items
+        within a spectral vector Euclidean distance of `3 * sigma` and
+        intensity above `vmin`.
+    vmin : float, optional
+        Signal intensity along `axis` below which spectra are excluded from
+        denoising.
+    dtype : dtype_like, optional
+        Data type of output arrays. Either float32 or float64.
+        The default is float64 unless the `signal` is float32.
+    num_threads : int, optional
+        Number of OpenMP threads to use for parallelization.
+        By default, multi-threading is disabled.
+        If zero, up to half of logical CPUs are used.
+        OpenMP may not be available on all platforms.
+
+    Returns
+    -------
+    ndarray
+        Denoised signal of `dtype`.
+        Spectra with integrated intensity below `vmin` are unchanged.
+
+    References
+    ----------
+    .. [2] Harman RC, Lang RT, Kercher EM, Leven P, and Spring BQ.
+       `Denoising multiplexed microscopy images in n-dimensional spectral space
+       <https://doi.org/10.1364/BOE.463979>`_.
+       *Biomed Opt Express*, 13(8): 4298-4309 (2022)
+
+    Notes
+    -----
+    This implementation is considered experimental. It is not validated
+    against the reference implementation and may not be practical with
+    real-world data. See discussion in `issue #142
+    <https://github.com/phasorpy/phasorpy/issues/142#issuecomment-2499421491>`_.
+
+    Examples
+    --------
+    Denoise a hyperspectral image with a Gaussian filter width of 0.1 in
+    spectral vector space using first and second harmonic:
+
+    >>> signal = numpy.random.randint(0, 255, (8, 16, 16))
+    >>> signal_filter_svd(signal, axis=0, sigma=0.1, harmonic=[1, 2])
+    array([[[...]]])
+
+    """
+    num_threads = number_threads(num_threads)
+
+    signal = numpy.asarray(signal)
+    if axis == -1 or axis == signal.ndim - 1:
+        axis = -1
+    else:
+        signal = numpy.moveaxis(signal, axis, -1)
+    shape = signal.shape
+    samples = shape[-1]
+
+    if harmonic is None:
+        harmonic = 1
+    harmonic, _ = parse_harmonic(harmonic, samples // 2)
+    num_harmonics = len(harmonic)
+
+    if vmin is None or vmin < 0.0:
+        vmin = 0.0
+
+    signal = numpy.ascontiguousarray(signal).reshape(-1, samples)
+    size = signal.shape[0]
+
+    if dtype is None:
+        if signal.dtype.char == 'f':
+            dtype = signal.dtype
+        else:
+            dtype = numpy.float64
+    dtype = numpy.dtype(dtype)
+    if dtype.char not in {'d', 'f'}:
+        raise ValueError('dtype is not floating point')
+
+    if spectral_vector is None:
+        sincos = numpy.empty((num_harmonics, samples, 2))
+        for i, h in enumerate(harmonic):
+            phase = numpy.linspace(
+                0,
+                h * math.pi * 2.0,
+                samples,
+                endpoint=False,
+                dtype=numpy.float64,
+            )
+            sincos[i, :, 0] = numpy.cos(phase)
+            sincos[i, :, 1] = numpy.sin(phase)
+        spectral_vector = numpy.zeros((size, num_harmonics * 2), dtype=dtype)
+
+        _phasor_from_signal_vector(
+            spectral_vector, signal, sincos, num_threads
+        )
+    else:
+        spectral_vector = numpy.ascontiguousarray(spectral_vector, dtype=dtype)
+        if spectral_vector.shape[:-1] != shape[:-1]:
+            raise ValueError('signal and spectral_vector shape mismatch')
+        spectral_vector = spectral_vector.reshape(
+            -1, spectral_vector.shape[-1]
+        )
+
+    if dtype == signal.dtype:
+        denoised = signal.copy()
+    else:
+        denoised = numpy.zeros(signal.shape, dtype=dtype)
+        denoised[:] = signal
+    integrated = numpy.zeros(size, dtype=dtype)
+    _signal_denoise_vector(
+        denoised, integrated, signal, spectral_vector, sigma, vmin, num_threads
+    )
+
+    denoised = denoised.reshape(shape)  # type: ignore[assignment]
+    if axis != -1:
+        denoised = numpy.moveaxis(denoised, -1, axis)
+    return denoised
+
+
+def signal_filter_ncpca(
+    signal: ArrayLike,
+    /,
+    n_components: int | float | str | None = 3,
+    *,
+    axis: int = -1,
+    dtype: DTypeLike | None = None,
+    **kwargs: Any,
+) -> NDArray[Any]:
+    """Return signal filtered by noise-corrected principal component analysis.
+
+    Apply noise-corrected Principal Component Analysis (NC-PCA) to denoise
+    signal containing shot noise. The signal is Poisson-normalized,
+    its dimensionality reduced by PCA with a specified number of components,
+    and then reconstructed according to [3]_.
+
+    Parameters
+    ----------
+    signal : array_like
+        Data containing Poisson noise to be filtered.
+        Must have at least 3 samples along the specified `axis`.
+    n_components : int, float, or str, optional, default: 3
+        Number of principal components to retain.
+        The default is 3, matching the reference implementation.
+        If None, all components are kept (no denoising).
+        If 'mle', use Minka's MLE to guess the dimension.
+        If 0 < n_components < 1 and svd_solver == 'full', select the number
+        of components such that the amount of variance that needs to be
+        explained is greater than the percentage specified by n_components.
+        See `sklearn.decomposition.PCA
+        <https://scikit-learn.org/stable/modules/generated/sklearn.decomposition.PCA.html>`_
+        for more details.
+    axis : int, optional, default: -1
+        Axis containing PCA features, for example, FLIM histogram bins.
+        The default is the last axis (-1).
+        Other axes are flattened and used as PCA samples.
+    dtype : dtype_like, optional
+        Data type of computation and output arrays. Either float32 or float64.
+        The default is float64 unless the input `signal` is float32.
+    **kwargs
+        Optional arguments passed to :py:class:`sklearn.decomposition.PCA`.
+
+    Returns
+    -------
+    ndarray
+        Denoised signal of specified `dtype`. Values may be negative.
+        Values, where the signal mean along `axis` is zero, are set to NaN.
+
+    Raises
+    ------
+    ValueError
+        If `dtype` is not a floating point type.
+        If `signal` has fewer than 3 samples along specified axis.
+        If `n_components` is invalid for the data size.
+
+    References
+    ----------
+    .. [3] Soltani S, Paulson J, Fong E, Mumenthaler, S, and Armani A.
+       `Denoising of fluorescence lifetime imaging data via principal
+       component analysis <https://doi.org/10.21203/rs.3.rs-7143126/v1>`_.
+       *Preprint*, (2025)
+
+    Notes
+    -----
+    Intensities of the reconstructed signal may be negative.
+    Hence, the phasor coordinates calculated from the reconstructed signal
+    may be outside the unit circle.
+    Consider thresholding low intensities for further analysis.
+
+    Examples
+    --------
+    Denoise FLIM data using 3 principal components:
+
+    >>> signal = numpy.random.poisson(100, (32, 32, 64))
+    >>> denoised = signal_filter_ncpca(signal, n_components=3)
+    >>> denoised.shape
+    (32, 32, 64)
+
+    """
+    from sklearn.decomposition import PCA
+
+    if (
+        dtype is None
+        and isinstance(signal, numpy.ndarray)
+        and signal.dtype == numpy.float32
+    ):
+        signal = signal.copy()
+        dtype = signal.dtype
+    else:
+        dtype = numpy.dtype(dtype)
+        if dtype.char not in {'f', 'd'}:
+            raise ValueError(f'{dtype=} is not a floating point type')
+        signal = numpy.asarray(signal, dtype=dtype, copy=True)
+
+    if axis == -1 or axis == signal.ndim - 1:
+        axis = -1
+    else:
+        signal = numpy.moveaxis(signal, axis, -1)
+
+    shape = signal.shape
+
+    if signal.size == 0:
+        raise ValueError('signal array is empty')
+    if signal.shape[-1] < 3:
+        raise ValueError(f'{signal.shape[-1]=} < 3')
+
+    # flatten sample dimensions
+    signal = signal.reshape(-1, shape[-1])
+
+    # poisson-normalize signal
+    scale = numpy.sqrt(numpy.nanmean(signal, axis=0, keepdims=True))
+    with numpy.errstate(divide='ignore', invalid='ignore'):
+        signal /= scale
+
+    # replace NaN and infinite values
+    mask = numpy.logical_or(
+        numpy.any(numpy.isnan(signal), axis=-1),
+        numpy.any(numpy.isinf(signal), axis=-1),
+    )
+    assert isinstance(signal, numpy.ndarray)  # for Mypy
+    signal[mask] = 0.0
+
+    # PCA transform and reconstruction with n_components
+    pca = PCA(n_components, **kwargs)
+    signal = pca.inverse_transform(pca.fit_transform(signal))
+    del pca
+
+    # restore original scale and shape
+    assert isinstance(signal, numpy.ndarray)  # for Mypy
+    signal[mask] = numpy.nan
+    signal *= scale
+    signal = signal.reshape(shape)
+    if axis != -1:
+        signal = numpy.moveaxis(signal, -1, axis)
+
+    return signal