python-microscopy 25.6.5__cp39-cp39-win_amd64.whl

PYME/Analysis/points/spherical_harmonics.py

@@ -0,0 +1,909 @@
+"""
+Estimate spherical harmonics from a point data set
+Initial fitting/conversions ripped 100% from David Baddeley / scipy
+"""
+from PYME.IO.image import ImageBounds
+import numpy as np
+from scipy.special import sph_harm
+from scipy import linalg
+from PYME.Analysis.points import coordinate_tools
+from scipy import optimize
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+def r_sph_harm(m, n, azimuth, zenith):
+    """
+    return real valued spherical harmonics. Uses the convention that m > 0 corresponds to the cosine terms, m < zero the
+    sine terms
+
+    Parameters
+    ----------
+    m : int
+        order of the spherical harmonic, |m| <= n
+    n : int
+        degree of the spherical harmonic, n >= 0
+
+    azimuth : ndarray
+        the azimuth angle in [0, 2pi]
+    zenith : ndarray
+        the elevation in [0, pi]
+
+    Returns
+    -------
+
+    """
+    if m > 0:
+        return (1. / np.sqrt(2) * (-1) ** m) * sph_harm(m, n, azimuth, zenith).real
+    elif m == 0:
+        return sph_harm(m, n, azimuth, zenith).real
+    else:
+        return (1. / np.sqrt(2) * (-1.0) ** m) * sph_harm(m, n, azimuth, zenith).imag
+
+
+def sphere_expansion(x, y, z, n_max=3):
+    """
+    Project coordinates onto spherical harmonics
+
+    Parameters
+    ----------
+    x : ndarray
+        x coordinates
+    y : ndarray
+        y coordinates
+    z : ndarray
+        z coordinates
+    n_max : int
+        Maximum order to calculate to
+
+    Returns
+    -------
+
+    modes : list of tuples
+        a list of the (m, n) modes projected onto
+    c : ndarray
+        the mode coefficients
+
+
+    """
+
+    azimuth, zenith, r = coordinate_tools.cartesian_to_spherical(x, y, z)
+
+    A = []
+    modes = []
+    for n in range(n_max + 1):
+        for m in range(-n, n + 1):
+            sp_mode = r_sph_harm(m, n, azimuth, zenith)
+            A.append(sp_mode)
+
+            modes.append((m, n))
+
+    A = np.vstack(A)
+
+    c = linalg.lstsq(A.T, r)[0]
+
+    return modes, c
+
+
+def sphere_expansion_clean(x, y, z, n_max=3, max_iters=2, tol_init=0.3):
+    """
+    Project coordinates onto spherical harmonics
+
+    Parameters
+    ----------
+    x : ndarray
+        x coordinates
+    y : ndarray
+        y coordinates
+    z : ndarray
+        z coordinates
+    n_max : int
+        Maximum order to calculate to
+    max_iters: int
+        number of fit iterations
+    tol_init: float
+        relative outlier tolerance. Used to ignore outliers in subsequent iterations
+
+    Returns
+    -------
+
+    modes : list of tuples
+        a list of the (m, n) modes projected onto
+    c : ndarray
+        the mode coefficients
+
+
+    """
+
+    azimuth, zenith, r = coordinate_tools.cartesian_to_spherical(x, y, z)
+
+    A = []
+    modes = []
+    for n in range(n_max + 1):
+        for m in range(-n, n + 1):
+            sp_mode = r_sph_harm(m, n, azimuth, zenith)
+            A.append(sp_mode)
+
+            modes.append((m, n))
+
+    A = np.vstack(A).T
+
+    tol = tol_init
+
+    c = linalg.lstsq(A, r)[0]
+
+    # recompute, discarding outliers
+    for i in range(max_iters):
+        pred = np.dot(A, c)
+        error = abs(r - pred) / r
+        mask = error < tol
+        # print mask.sum(), len(mask)
+
+        c, summed_residuals, rank, singular_values = linalg.lstsq(A[mask, :], r[mask])
+        tol /= 2
+
+    return modes, c, summed_residuals
+
+
+AXES = np.stack([[1., 0., 0.], [0., 1., 0.], [0., 0., 1.]], axis=1)
+
+
+def reconstruct_shell(modes, coeffs, azimuth, zenith):
+    r = 0
+    # TODO save some compute by using a single call for + and - m modes
+    cache = {}
+    for (m, n), c in zip(modes, coeffs):
+        if m == 0:
+            r += c*sph_harm(m, n, azimuth, zenith).real
+        else:
+            k = (abs(m), n) # sph_harm returns both +m and -m as real and imag parts, make sure we don't evaluate twice
+            try:
+                s = cache[k]
+            except KeyError:
+                s = sph_harm(m,n,azimuth, zenith)
+                cache[k] = s
+
+            if m > 0:
+                r += (c * ((1. / np.sqrt(2) * (-1) ** m) *s.real)).astype(azimuth.dtype)
+            else:
+                r += (c * ((1. / np.sqrt(2) * (-1) ** m) *s.imag)).astype(azimuth.dtype)
+
+    return r
+
+
+def scaled_shell_from_hdf(hdf_file, table_name='harmonic_shell'):
+    """
+
+    Parameters
+    ----------
+    hdf_file : str or tables.file.File
+        path to hdf file or opened hdf file
+    table_name : str
+        name of the table containing the spherical harmonic expansion information
+
+    Returns
+    -------
+    shell : ScaledShell
+        see nucleus.spherical_harmonics.shell_tools.ScaledShell
+
+    """
+    from PYME.IO.MetaDataHandler import HDFMDHandler
+    from PYME.IO import tabular
+    import tables
+    try:
+        opened_hdf_file = tables.open_file(hdf_file, 'r')
+    except TypeError:
+        opened_hdf_file = hdf_file
+    
+    shell_table = tabular.HDFSource(opened_hdf_file, table_name)
+    shell_table.mdh = HDFMDHandler(opened_hdf_file)
+
+    return ScaledShell.from_tabular(shell_table)
+
+def generate_icosahedron():
+    """
+    Generate an icosahedron in spherical coordinates.
+    
+    Returns
+    -------
+        azimuth : np.array
+            Length 12, azimuth of icosahedron vertices [0, 2*pi].
+        zenith : np.array
+            Length 12, zenith of icosahedron vertices [0, pi].
+        faces : np.array
+            20 x 3, counterclockwise triangles denoting icosahedron faces.
+    """
+    t = np.arctan(0.5)+np.pi/2
+    zenith = np.hstack([0, np.array(5*[np.pi-t,t]).ravel(), np.pi])  # [0, pi]
+    azimuth = np.hstack([0, np.arange(0, 2*np.pi, np.pi/5), 0])   # [0, 2*pi]
+    idxs = np.arange(len(zenith))[1:-1]
+    upper_middle_strip = np.vstack([[v0,v2,v1] for v0,v1,v2 in 
+                                    zip(idxs[::2],idxs[1::2],np.roll(idxs[::2],-1))])
+    lower_middle_strip = np.vstack([[v0,v2,v1] for v0,v1,v2 in 
+                                    zip(np.roll(idxs[::2],-1),idxs[1::2],np.roll(idxs[1::2],-1))])
+    upper_cap = np.vstack([[v0,v2,v1] for v0,v1,v2 in 
+                        zip(np.zeros(5),idxs[::2],np.roll(idxs[::2],-1))])
+    lower_cap = np.vstack([[v0,v2,v1] for v0,v1,v2 in 
+                        zip(11*np.ones(5),np.roll(idxs[1::2],-1),idxs[1::2])])
+    faces = np.vstack([upper_cap, upper_middle_strip, lower_middle_strip, lower_cap]).astype(int)
+    
+    return azimuth, zenith, faces
+
+def icosahedron_mesh(n_subdivision=1):
+    """
+    Return an icosahedron subdivided n_subdivision times. This provides a
+    quasi-regular sampling of the unit sphere. ::
+        
+            v0               v0
+            /  \             /  \
+           /    \    ==>   v3----v5
+          /      \         / \  / \
+         v1------v2       v1--v4--v2
+
+    Returns
+    -------
+        azimuth : np.array
+            Azimuth of mesh vertices [0, 2*pi].
+        zenith : np.array
+            Zenith of mesh vertices [0, pi].
+        faces : np.array
+            20 x 3, counterclockwise triangles denoting mesh faces.
+    """
+    
+    # First, generate the icosahedron
+    azimuth, zenith, faces = generate_icosahedron()
+    
+    if n_subdivision < 1:
+        # We don't need to do anything
+        return azimuth, zenith, faces
+    
+    # Convert pole azimuth to np.nan so we can ignore pole azimuth (irrelevant, detrimental)
+    azimuth[0], azimuth[-1] = np.nan, np.nan
+    
+    # Make azimuth and zenith complex numbers (to deal with phase wrapping)
+    azimuth, zenith = np.exp(1j*azimuth), np.exp(1j*zenith)
+    
+    for k in range(n_subdivision):
+        # Average each edge to get new vertex at the center of each face v0,v1,v2
+        az_v3 = np.nanmean(azimuth[faces[:,[0,1]]],axis=1)
+        az_v4 = np.nanmean(azimuth[faces[:,[1,2]]],axis=1)
+        az_v5 = np.nanmean(azimuth[faces[:,[2,0]]],axis=1)
+        new_az = np.hstack([az_v3, az_v4, az_v5])
+        ze_v3 = np.nanmean(zenith[faces[:,[0,1]]],axis=1)
+        ze_v4 = np.nanmean(zenith[faces[:,[1,2]]],axis=1)
+        ze_v5 = np.nanmean(zenith[faces[:,[2,0]]],axis=1)
+        new_ze = np.hstack([ze_v3, ze_v4, ze_v5])
+        
+        # Create new vertices, eliminating duplicates (there should be two of each 
+        # new vertex, one from each face sharing an edge)
+        new_v, new_idxs = np.unique(np.vstack([new_az,new_ze]).T, axis=0, return_inverse=True)
+        new_idxs += len(azimuth)
+        new_idxs = new_idxs.reshape(3,-1).T  # v3 = new_idxs[:,0], v4 = new_idxs[:,1], v5 = new_idxs[:,2]
+        
+        # Create new faces (4 faces per old face)
+        f0 = np.vstack([faces[:,0], new_idxs[:,0], new_idxs[:,2]]).T
+        f1 = np.vstack([faces[:,1], new_idxs[:,1], new_idxs[:,0]]).T
+        f2 = np.vstack([faces[:,2], new_idxs[:,2], new_idxs[:,1]]).T
+        faces = np.vstack([f0,f1,f2,new_idxs]).astype(int)
+        
+        # Append the new vertices to azimuth
+        azimuth = np.hstack([azimuth, new_v[:,0]])
+        zenith = np.hstack([zenith, new_v[:,1]])
+    
+    # Restore poles to azimuth 0 for future conversion to real coordinates
+    azimuth[np.isnan(azimuth)] = 0
+    
+    # Back to angles with you
+    azimuth, zenith = np.angle(azimuth), np.angle(zenith)
+    azimuth[azimuth<0] = azimuth[azimuth<0]+2*np.pi  # wrap to [0, 2*pi]
+    zenith = np.abs(zenith) # wrap to [0, pi]
+        
+    return azimuth, zenith, faces
+
+
+class ScaledShell(object):
+    data_type = [
+        ('modes', '<2i4'),
+        ('coefficients', '<f4'),
+    ]
+
+    def __init__(self, sampling_fraction=1.):
+        self.sampling_fraction = sampling_fraction
+        self.modes = None
+        self.coefficients = None
+
+        self.x, self.y, self.z, = None, None, None
+        self.x0, self.y0, self.z0 = None, None, None
+        self.x_c, self.y_c, self.z_c, = None, None, None
+        # note that all scalings will be centered
+        self.x_cs, self.y_cs, self.z_cs, = None, None, None
+
+        self.standard_deviations, self.principal_axes = None, None
+        self.scaling_factors = None
+        self._table_repr = None
+
+    @property
+    def table_representation(self):
+        if self._table_repr is None:
+            self._update_table_repr()
+
+        return self._table_repr
+
+    def _update_table_repr(self):
+        from PYME.IO import tabular
+        self._table_repr = tabular.scalar_column_source(
+            standard_deviations=self.standard_deviations,
+            scaling_factors = self.scaling_factors,
+            principal_axes = self.principal_axes,
+            summed_residuals = self._summed_residuals,
+            n_points_fitted = len(self.x),
+            x0 = self.x0,
+            y0 = self.y0,
+            z0 = self.z0,
+            sampling_fraction = self.sampling_fraction,
+            modes = np.asarray(self.modes, dtype='i4'), 
+            coefficients=np.asarray(self.coefficients, dtype='f4'),
+        ) 
+    
+    def to_recarray(self, keys=None):
+        """
+
+        Pretend we are a PYME.IO.tabular type
+
+        Parameters
+        ----------
+        keys : None
+            Ignored for this contrived function
+
+        Returns
+        -------
+        numpy recarray version of self
+
+        """
+
+        return self.table_representation.to_recarray(self, keys=keys)
+
+    def to_hdf(self, *args, **kwargs):
+        return self.table_representation.to_hdf(*args, **kwargs)
+
+
+    @staticmethod
+    def from_tabular(shell_table, index=0):
+        shell = ScaledShell()
+
+        if len(shell_table.keys() == 2):
+            # old legacy format which stores data in metadata
+        
+            shell.standard_deviations = np.asarray(shell_table.mdh['spherical_harmonic_shell.standard_deviations'])
+            shell.scaling_factors = np.asarray(shell_table.mdh['spherical_harmonic_shell.scaling_factors'])
+            shell.principal_axes = np.asarray(shell_table.mdh['spherical_harmonic_shell.principal_axes'])
+
+            shell.x0 = shell_table.mdh['spherical_harmonic_shell.x0']
+            shell.y0 = shell_table.mdh['spherical_harmonic_shell.y0']
+            shell.z0 = shell_table.mdh['spherical_harmonic_shell.z0']
+
+            shell._summed_residuals = shell_table.mdh['spherical_harmonic_shell.summed_residuals']
+            shell.sampling_fraction = shell_table.mdh['spherical_harmonic_shell.sampling_fraction']
+            
+            shell._set_coefficients(shell_table['modes'], shell_table['coefficients'])
+
+        else:
+            #new format which stores everything in one table
+            shell.standard_deviations = shell_table['standard_deviations'][index]
+            shell.scaling_factors = shell_table['scaling_factors'][index]
+            shell.principal_axes = shell_table['principal_axes'][index]
+
+            shell.x0 = shell_table['x0'][index]
+            shell.y0 = shell_table['y0'][index]
+            shell.z0 = shell_table['z0'][index]
+
+            shell._summed_residuals = shell_table['summed_residuals'][index]
+            shell.sampling_fraction = shell_table['sampling_fraction'][index]
+            
+            shell._set_coefficients(shell_table['modes'][index], shell_table['coefficients'][index])
+
+        return shell
+
+
+    def _set_coefficients(self, modes, coefficients):
+        assert len(modes) == len(coefficients)
+        self.modes = modes
+        self.coefficients = coefficients
+
+    def set_fitting_points(self, x, y, z):
+        from PYME.IO.image import ImageBounds
+
+        assert (x.shape == y.shape) and (y.shape == z.shape)
+        self.x, self.y, self.z = x.astype(np.float32, copy=True), y.astype(np.float32, copy=True), z.astype(np.float32, copy=True)
+        
+        #does it really make sense to use ImageBounds here??
+        self._fitting_point_bounds = ImageBounds(self.x.min(), self.y.min(),
+                                                 self.x.max(), self.y.max(),
+                                                 self.z.min(), self.z.max())
+        self.x0, self.y0, self.z0 = self.x.mean(), self.y.mean(), self.z.mean()
+
+        self.x_c, self.y_c, self.z_c = self.x - self.x0, self.y - self.y0, self.z - self.z0
+
+        self._scale_fitting_points()
+
+    def _scale_fitting_points(self):
+        self.standard_deviations, self.principal_axes = coordinate_tools.find_principle_axes(self.x_c, self.y_c, self.z_c,
+                                                                                             sample_fraction=self.sampling_fraction)
+        self.scaling_factors = np.max(self.standard_deviations) /(self.standard_deviations)
+        self.x_cs, self.y_cs, self.z_cs, = coordinate_tools.scaled_projection(self.x_c, self.y_c, self.z_c,
+                                                                              self.scaling_factors, self.principal_axes)
+
+    def shell_coordinates(self, points):
+        """Scale query points, projecting them onto the basis used in shell-
+        fitting. Return in scaled spherical coordinates
+
+        Parameters
+        ----------
+        points : 3-tuple
+            x, y, z positions
+
+        Returns
+        -------
+        azimuth, zenith, r
+            scaled spherical coordinates of input points
+        """
+        x, y, z = points
+
+        # scale the query points and convert them to spherical
+        x_qs, y_qs, z_qs = coordinate_tools.scaled_projection(np.atleast_1d(x - self.x0), np.atleast_1d(y - self.y0),
+                                                              np.atleast_1d(z - self.z0), self.scaling_factors,
+                                                              self.principal_axes)
+        
+        azimuth_qs, zenith_qs, r_qs = coordinate_tools.cartesian_to_spherical(x_qs, y_qs, z_qs)
+        return azimuth_qs, zenith_qs, r_qs
+
+    def get_fitted_shell(self, azimuth, zenith):
+        r_scaled = reconstruct_shell(self.modes, self.coefficients, azimuth, zenith)
+
+        return self._to_cartesian(azimuth, zenith, r_scaled)
+    
+    def _to_cartesian(self, azimuth, zenith, r_scaled):
+        x_scaled, y_scaled, z_scaled = coordinate_tools.spherical_to_cartesian(azimuth, zenith, r_scaled)
+        # need to scale things "down" since they were scaled "up" in the fit
+        # scaling_factors = 1. / self.scaling_factors
+
+        scaled_axes = (self.principal_axes / self.scaling_factors[:, None]).astype(azimuth.dtype, copy=False)
+
+        coords = x_scaled.ravel()[:, None] * scaled_axes[0, :] + y_scaled.ravel()[:, None] * scaled_axes[1,
+                                                                                             :] + z_scaled.ravel()[:,
+                                                                                                  None] * scaled_axes[2,
+                                                                                                          :]
+        x, y, z = coords.T
+
+        return x.reshape(x_scaled.shape) + self.x0, y.reshape(y_scaled.shape) + self.y0, z.reshape(
+            z_scaled.shape) + self.z0
+
+
+    def fit_shell(self, max_n_mode=3, max_iterations=2, tol_init=0.3):
+        modes, coefficients, summed_residuals = sphere_expansion_clean(self.x_cs, self.y_cs, self.z_cs, max_n_mode,
+                                                                       max_iterations, tol_init)
+        self._set_coefficients(modes, coefficients)
+        self._summed_residuals = summed_residuals
+
+    def check_inside(self, x=None, y=None, z=None):
+        if x is None:
+            xcs, ycs, zcs = self.x_cs, self.y_cs, self.z_cs
+        else:
+            xcs, ycs, zcs = coordinate_tools.scaled_projection(x - self.x0, y - self.y0, z - self.z0, self.scaling_factors,
+                                                               self.principal_axes)
+
+        azimuth, zenith, rcs = coordinate_tools.cartesian_to_spherical(xcs, ycs, zcs)
+        r_cs_shell = reconstruct_shell(self.modes, self.coefficients, azimuth, zenith)
+        return rcs < r_cs_shell
+
+    def _visualize_shell(self, d_zenith=0.1, points=None):
+        try:
+            from mayavi import mlab
+        except(ImportError):
+            raise ImportError('Could not import mayavi.mlab.\
+             Please make sure mayavi is installed to display fitted shell')
+
+        if not points:
+            x, y, z = self.x, self.y, self.z
+        else:
+            x, y, z = points
+        zenith, azimuth = np.mgrid[0:(np.pi + d_zenith):d_zenith, 0:(2 * np.pi + d_zenith):d_zenith]
+
+        xs, ys, zs = self.get_fitted_shell(azimuth, zenith)
+
+        mlab.figure()
+        mlab.mesh(xs, ys, zs)
+        mlab.points3d(x, y, z, mode='point')
+
+    def get_mesh_vertices_faces(self, d_zenith=0.1):
+        """Compute vertices and faces to pass to 
+        PYME.experimental._triangle_mesh.TriangleMesh.
+
+        Note this will be non-manifold because of cuts at zenith=0, azimuth=0.
+
+        Parameters
+        ----------
+        d_zenith : float, optional
+            zenith step size for generating vector in plane of the shell [radians], by default 0.1
+        """
+
+        # Convert d_zenith to closest subdivision >= d_zenith (limited control due to icosahedon subdivision)
+        # Mean initial d_zenith is 2*pi/5, so solve for (2*pi/10)^n_subdivision = d_zenith
+        n_subdivision = int(np.ceil(np.log(d_zenith)/(np.log(np.pi)-np.log(5))))
+
+        # Quasi-regular sample points on a unit sphere with icosahedron subdivision
+        azimuth, zenith, faces = icosahedron_mesh(n_subdivision)
+        # Map points onto shell radius 
+        xs, ys, zs = self.get_fitted_shell(azimuth, zenith)
+
+        # Create vertices 
+        vertices = np.vstack([xs.ravel(), ys.ravel(), zs.ravel()]).T   # Row-major ravel
+
+        return vertices.astype(np.float32), faces.astype(np.int32)
+
+    # def _visualize_scaled(self):
+    #     from mayavi import mlab
+    #     visualize_shell(self.modes, self.coefficients)#, scaling_factors=self.standard_deviations,
+    #                     # scaling_axes=self.principal_axes)
+    #     mlab.points3d(self.x_cs, self.y_cs, self.z_cs, mode='point')
+
+    # def _visualize(self):
+    #     from mayavi import mlab
+    #     visualize_shell(self.modes, self.coefficients, scaling_factors=1./self.scaling_factors,
+    #                     scaling_axes=self.principal_axes)
+    #     mlab.points3d(self.x_c, self.y_c, self.z_c, mode='point')
+
+    def distance_to_shell(self, query_points, d_angles=0.1):
+        """
+
+        Parameters
+        ----------
+        query_points : list-like of ndarrays
+            Arrays of positions to query (in cartesian coordinates), i.e. [np.array(x), np.array(y), np.array(z)]
+        d_angles : float
+            Sets the step size in radians of zenith and azimuth arrays used in reconstructing the spherical harmonic shell
+
+        Returns
+        -------
+        min_distance : float
+            minimum distance from query points (i.e. input coordinate) to the spherical harmonic surface
+        closest_points_on_surface : tuple of floats
+            returns the position in cartesian coordinates of the point on the surface closest to the input 'position'
+
+        """
+        x, y, z = query_points
+        x, y, z = np.atleast_1d(x), np.atleast_1d(y), np.atleast_1d(z)
+        n_points = len(x)
+        zenith, azimuth = np.mgrid[0:(np.pi + d_angles):d_angles, 0:(2 * np.pi + d_angles):d_angles]
+
+        x_shell, y_shell, z_shell = self.get_fitted_shell(azimuth.astype(np.float32, copy=False),
+                                                          zenith.astype(np.float32, copy=False))
+        # calculate the distance between all our points and the shell
+        dist = np.sqrt(
+            (x - x_shell[:, :, None]) ** 2 + (y - y_shell[:, :, None]) ** 2 + ((z - z_shell[:, :, None]) ** 2))
+
+        # unfortunately cannot currently specify two axes for numpy.argmin, so we'll have to flatten the first two dims
+        n_shell_coords = dist.shape[0] * dist.shape[1]
+        dist_flat = dist.reshape((n_shell_coords, n_points))
+        min_ind = np.argmin(dist_flat, axis=0)
+
+        p_ind = range(n_points)
+        return dist_flat[min_ind[p_ind], p_ind], (x_shell.reshape(n_shell_coords)[min_ind],
+                                                  y_shell.reshape(n_shell_coords)[min_ind],
+                                                  z_shell.reshape(n_shell_coords)[min_ind])
+    
+    def radial_distance_to_shell(self, points):
+        """ finds distance to shell along a vector from the centre of the shell to a point.
+        
+        Notes
+        -----
+        
+        This is not a geometric/euclidean distance, but should be a good approximation close to the shell (or far away).
+        Over mid-range distances, it will give potentially very distorted distances. It is, however, very much faster and
+        can safely be used in applications where strict scaling is not important - e.g. for a SDF representation of the
+        surface.
+        
+        """
+        # project points into scaled-space spherical coordinates
+        azimuth_qs, zenith_qs, r_qs = self.shell_coordinates(points)
+
+        # get scaled shell radius at those angles
+        r_shell = reconstruct_shell(self.modes, self.coefficients, azimuth_qs, zenith_qs)
+        pts = self._to_cartesian(azimuth_qs, zenith_qs, r_shell)
+
+        #print(r_qs - r_shell)
+
+        # return the (scaled space) difference
+        #d = r_qs - r_shell
+        d_ = pts - points
+
+        s = 2.0*(r_qs> r_shell) - 1.0
+        #print(s)
+
+        d = s*np.sqrt((d_*d_).sum(0))
+        
+        #scale distance TODO - actually scale to make slightly closer to being euclidean
+        return d
+        
+
+    def approximate_normal(self, x, y, z, d_azimuth=1e-6, d_zenith=1e-6, return_orthogonal_vectors=False):
+        """
+
+        Numerically approximate a vector(s) normal to the spherical harmonic shell at the query point(s).
+
+        For input point(s), scale and convert to spherical coordinates, shift by +/- d_azimuth and d_zenith to get
+        'phantom' points in the plane tangent to the spherical harmonic expansion on either side of the query point(s).
+        Scale back, convert to cartesian, make vectors from the phantom points (which are by definition not parallel)
+        and cross them to get a vector perpindicular to the plane.
+
+        Parameters
+        ----------
+        x : ndarray, float
+            cartesian x location of point(s) on the surface to calculate the normal at
+        y : ndarray, float
+            cartesian y location of point(s) on the surface to calculate the normal at
+        z : ndarray, float
+            cartesian z location of point(s) on the surface to calculate the normal at
+        d_azimuth : float
+            azimuth step size for generating vector in plane of the shell [radians]
+        d_zenith : float
+            zenith step size for generating vector in plane of the shell [radians]
+
+        Returns
+        -------
+        normal_vector : ndarray
+            cartesian unit vector(s) normal to query point(s). size (len(x), 3)
+        orth0 : ndarray
+            cartesian unit vector(s) in the plane of the spherical harmonic shell at the query point(s), and
+            perpendicular to normal_vector
+        orth1 : ndarray
+            cartesian unit vector(s) orthogonal to normal_vector and orth0
+
+        """
+        # scale the query points and convert them to spherical
+        x_qs, y_qs, z_qs = coordinate_tools.scaled_projection(np.atleast_1d(x - self.x0), np.atleast_1d(y - self.y0),
+                                                              np.atleast_1d(z - self.z0), self.scaling_factors,
+                                                              self.principal_axes)
+        azimuth, zenith, r = coordinate_tools.cartesian_to_spherical(x_qs, y_qs, z_qs)
+
+        # get scaled shell radius at +/- points for azimuthal and zenith shifts
+        azimuths = np.array([azimuth - d_azimuth, azimuth + d_azimuth, azimuth, azimuth])
+        zeniths = np.array([zenith, zenith, zenith - d_zenith, zenith + d_zenith])
+        r_scaled = reconstruct_shell(self.modes, self.coefficients, azimuths, zeniths)
+
+        # convert shifted points to cartesian and scale back. shape = (4, #points)
+        x_scaled, y_scaled, z_scaled = coordinate_tools.spherical_to_cartesian(azimuths, zeniths, r_scaled)
+        # scale things "down" since they were scaled "up" in the fit
+        scaled_axes = self.principal_axes / self.scaling_factors[:, None]
+        coords = x_scaled.ravel()[:, None] * scaled_axes[0, :] + \
+                 y_scaled.ravel()[:, None] * scaled_axes[1, :] + \
+                 z_scaled.ravel()[:, None] * scaled_axes[2, :]
+        x_p, y_p, z_p = coords.T
+        # skip adding x0, y0, z0 back on, since we'll subtract it off in a second
+        x_p, y_p, z_p = x_p.reshape(x_scaled.shape), y_p.reshape(y_scaled.shape), z_p.reshape(z_scaled.shape)
+
+        # make two vectors in the plane centered at the query point
+        v0 = np.array([x_p[1] - x_p[0], y_p[1] - y_p[0], z_p[1] - z_p[0]])
+        v1 = np.array([x_p[3] - x_p[2], y_p[3] - y_p[2], z_p[3] - z_p[2]])
+        if not np.any(v0) or not np.any(v1):
+            raise RuntimeWarning('failed to generate two vectors in the plane - likely precision error in sph -> cart')
+        # cross them to get a normal vector NOTE - direction could be negative of true normal
+        normal = np.cross(v0, v1, axis=0)
+        # return as unit vector(s) along each row
+        normal = np.atleast_2d(normal / np.linalg.norm(normal, axis=0)).T
+        # make sure normals point outwards, by dotting it with the vector to the point on the shell from the center
+        points = np.stack([np.atleast_1d(x - self.x0), np.atleast_1d(y - self.y0), np.atleast_1d(z - self.z0)]).T
+        outwards = np.array([np.dot(normal[ind], points[ind]) > 0 for ind in range(normal.shape[0])])
+        normal[~outwards, :] *= -1
+        if np.isnan(normal).any():
+            raise RuntimeError('Failed to calculate normal vector')
+        if return_orthogonal_vectors:
+            orth0 = np.atleast_2d(v0 / np.linalg.norm(v0, axis=0)).T
+            # v0 and v1 are both in a plane perpendicular to normal, but not strictly orthogonal to each other
+            orth1 = np.cross(normal, orth0, axis=1)  # replace v1 with a unit vector orth. to both normal and v0
+            return normal.squeeze(), orth0.squeeze(), orth1.squeeze()
+        return normal.squeeze()
+
+    def _distance_error(self, parameterized_distance, vector, starting_point):
+        """
+
+        Calculate the error in scaled space between the shell and the point reached traveling a specified distance(s)
+        along the input vector from the input starting position.
+
+        This function is to be minimized by a solver. Note that we don't actually have to calculate the distance in
+        normal space, since minimizing in the scale space is equivalent.
+
+        Parameters
+        ----------
+        parameterized_distance : float
+            distance along vector to travel, units of nm, unscaled
+        vector : ndarray
+            Length three, vector in (unscaled) cartesian coordinates along which to travel
+        starting_point : ndarray or list
+            Length three, point in (unscaled) cartesian space from which to start traveling along 'vector'
+
+        Returns
+        -------
+
+        """
+        x, y, z = [parameterized_distance * np.atleast_2d(vector)[:, ind] + np.atleast_2d(starting_point)[:, ind] for
+                   ind in range(3)]
+        # scale the query points and convert them to spherical
+        x_qs, y_qs, z_qs = coordinate_tools.scaled_projection(np.atleast_1d(x - self.x0), np.atleast_1d(y - self.y0),
+                                                              np.atleast_1d(z - self.z0), self.scaling_factors,
+                                                              self.principal_axes)
+        azimuth_qs, zenith_qs, r_qs = coordinate_tools.cartesian_to_spherical(x_qs, y_qs, z_qs)
+
+        # get scaled shell radius at those angles
+        r_shell = reconstruct_shell(self.modes, self.coefficients, azimuth_qs, zenith_qs)
+
+        # return the (scaled space) difference
+        return r_qs - r_shell
+
+    def distance_to_shell_along_vector_from_point(self, vector, starting_point, guess=None):
+        """
+
+        Calculate the distance to the shell along a given direction, from a given point.
+
+        Parameters
+        ----------
+        vector : list-like
+            cartesian vector indicating direction to query for proximity to shell
+        starting_point : list-like
+            cartesian position from which to start traveling along input vector when calculating shell proximity
+        guess_distances : array, float
+            initial guess for distance solver. See self._distance_error()
+
+        Returns
+        -------
+
+        """
+
+        if guess is None:
+            guess = self._find_guess_for_distance_to_shell_along_vector_from_point(vector, starting_point)
+
+        (res, cov_x, info_dict, mesg, res_code) = optimize.leastsq(self._distance_error, guess,
+                                                                   args=(vector, starting_point),
+                                                                   full_output=1)
+        return res
+
+    def _find_guess_for_distance_to_shell_along_vector_from_point(self, vector, starting_point, guess_distances=None):
+        """
+
+        Calculate the distance to the shell along a given direction, from a given point.
+
+        Parameters
+        ----------
+        vector : list-like
+            cartesian vector indicating direction to query for proximity to shell
+        starting_point : list-like
+            cartesian position from which to start traveling along input vector when calculating shell proximity
+        guess_distances : array, float
+            initial guess for distance solver. See self._distance_error()
+
+        Returns
+        -------
+
+        """
+        guess_distances = np.arange(0., 1000., 100.) if guess_distances is None else guess_distances
+        errors = np.zeros_like(guess_distances)
+        # guess = guess_distances[np.argmin(np.abs(self._distance_error(guess_distances, starting_point, vector)))]
+        for ind, query in enumerate(guess_distances):
+            errors[ind] = self._distance_error(query, vector, starting_point)
+
+        return guess_distances[np.argmin(np.abs(errors))]
+    
+    def approximate_image_bounds(self, d_zenith=0.1, d_azimuth=0.1):
+        from PYME.IO.image import ImageBounds
+        #does it really make sense to use ImageBounds here??
+        
+        zenith, azimuth = np.mgrid[0:(np.pi + d_zenith):d_zenith,
+                                   0:(2 * np.pi + d_azimuth):d_azimuth]
+
+        x_shell, y_shell, z_shell = self.get_fitted_shell(azimuth, zenith)
+
+        return ImageBounds(x_shell.min(), y_shell.min(), 
+                          x_shell.max(), y_shell.max(),
+                          z_shell.min(), z_shell.max())
+
+    def uniform_random_radial_density(self, n_radial_bins=20, batch_size=100000, target_sampling_nm=75.):
+        """
+        Estimate a radial density histogram for uniform spatial sampling within the shell. Performs calculation in batches to give bounded memory usage.
+
+        Parameters
+        ----------
+
+        n_radial_bins : int
+            Number of radial bins to divide the radius into
+        batch_size : int
+            number of test points to generate in one batch. Note that the default (100000) might be a bit conservative
+        target_sampling_nm: float
+            desired sampling density of combined batches. Used in conjuction with batch_size to determine the number of batches required.
+        """
+
+        bounds = self.approximate_image_bounds()
+
+        w = bounds.x1 - bounds.x0 #+ 2*target_sampling_nm
+        h = bounds.y1 - bounds.y0 #+ 2*target_sampling_nm
+        d = bounds.z1 - bounds.z0 #+ 2*target_sampling_nm
+
+        # to ensure spatial uniformity, simulate in a cube with an edge length equal to the longest bound.
+        # this means we chuck away quite a few points at the rejection step, but is logically simple
+        cube_size =max([w, h, d])
+
+        # calculate number of batches to give uniform random sampling with a desired spacing,
+        # whilst not exceeding the memory requirements dictated by a single batch 
+        n_batches = int(((cube_size/target_sampling_nm)**3)/batch_size)
+        logger.debug('Using %d batches for density estimation' % n_batches)
+
+        # calculate where to put our edges
+        bin_edges = np.linspace(0, 1, n_radial_bins)
+
+        #pre-allocate and output array
+        counts = np.zeros(len(bin_edges)-1, 'i4')
+
+        for i in range(n_batches):
+            # simulate batch_size uniformly distributed points within a bounding cube
+            x = np.random.uniform(bounds.x0, bounds.x0 + cube_size, batch_size)
+            y = np.random.uniform(bounds.y0, bounds.y0 + cube_size, batch_size)
+            z = np.random.uniform(bounds.z0, bounds.z0 + cube_size, batch_size)
+
+            # if shell_coordinates and/or reconstruct_shell prove expensive, we could clip to the rectangular (non-cube) bounding box here
+            # before calculating shell coordinates. Omitting for now as this is likely to be premature optimisation (and the additionaly indexing
+            # might well prove more expensive than just calculating the shell co-ordinates)
+
+            # convert these to polar, shell referenced coordinates
+            azi, zen, r = self.shell_coordinates((x, y, z))
+            r_shell = reconstruct_shell(self.modes,self.coefficients,azi, zen)
+
+            # find which points are inside the shell
+            inside = r < r_shell
+            # reject all points outside the shell
+            r_norm = r[inside]/r_shell[inside]
+
+            counts += np.histogram(r_norm, bin_edges)[0]
+
+        return bin_edges, counts
+
+
+
+class SHShell(ScaledShell):
+    '''
+    Initial work on a replacement interface for ScaledShell
+    
+    Goals:
+    
+    - can be constructed directly as well as / instead of being fit
+    - fitting is a single function call (rather than 3)
+    - easily serialised (TODO)
+    - does not own/keep a copy of data points (to facilitate serialisation)
+    - when directly constructed, can used to represent synthetic nuclei etc ... whilst retaining distance
+      evaluation functions
+    
+    '''
+    def __init__(self, centre=(0,0,0), principle_axes=((1,0,0), (0,1,0), (0,0,1)), axis_scaling=(1.,1.,1.), modes=((0,0),), coefficients=(1,)):
+        self.x0, self.y0, self.z0 = centre
+        self.principal_axes = np.array(principle_axes)
+        self.axis_scaling = np.array(axis_scaling)
+        self.modes = modes
+        self.coefficients = np.array(coefficients)
+        
+    def fit(self, points, n_max=3, n_iters=2, tol=0.3, principle_axis_sampling=1.0):
+        x, y, z = points
+        self.x0, self.y0, self.z0 = x.mean(), y.mean(), z.mean()
+        x_c, y_c, z_c = x - self.x0, y - self.y0, z - self.z0
+        
+        sig, self.principal_axes = coordinate_tools.find_principle_axes(x_c, y_c, z_c, sample_fraction=principle_axis_sampling)
+        self.axis_scaling = sig/np.max(sig)
+
+        x_cs, y_cs, z_cs, = coordinate_tools.scaled_projection(x_c, y_c, z_c,self.axis_scaling, self.principal_axes)
+
+        self.modes, self.coefficients, _ = sphere_expansion_clean(x_cs, y_cs, z_cs, n_max, n_iters, tol)
+        
+    @property
+    def scaling_factors(self):
+        import warnings
+        warnings.warn(DeprecationWarning('use .axis_scaling instead'))
+        return self.axis_scaling
+