scripts/test_fit_ellipses_anisotropy.py

"""Regression tests for the A1/A2 ellipse-fitting bugfix in
``MedVision_BenchmarkPlannerBiometry_fromSeg.__fit_ellipses``.

The biometry planner fits the lesion ellipse in *physical* (real-world) space so
the major/minor axes are physically meaningful. Two bugs broke that under
anisotropic in-plane spacing:

  A1 - the contour was scaled with transposed pixel spacing (cv2 points are
       ``(x=dim1, y=dim0)`` but ``pixel_sizes`` is ``(dim0, dim1)``);
  A2 - major vs. minor was decided by *pixel* length while reported in *mm*.

See ``doc/ellipse-fitting-image-vs-real-space.md`` for the linear algebra.

This is a plain-``assert`` script (the repo has no test framework); run with::

    python scripts/test_fit_ellipses_anisotropy.py

Exit code 0 = all tests passed.
"""

import os
import sys

import numpy as np

sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "src"))

from medvision_ds.utils.benchmark_planner import (  # noqa: E402
    MedVision_BenchmarkPlannerBiometry_fromSeg as Planner,
)

# private method name after Python name-mangling
_FIT = "_MedVision_BenchmarkPlannerBiometry_fromSeg__fit_ellipses"


def _make_planner():
    """Instantiate without the heavy DB-backed ``__init__``; ``__fit_ellipses``
    only needs the two bbox-scale attributes. The scales are set wide/narrow so
    the ``all_within`` acceptance filter never drops a valid landmark, isolating
    the fitting logic under test."""
    obj = Planner.__new__(Planner)
    obj.enlarged_bbox_scale = 5.0
    obj.shrunk_bbox_scale = 0.05
    return obj


def _rasterize_physical_ellipse(shape, center_idx, s0, s1, A, B, phi_deg):
    """Binary mask whose TRUE physical fit is an ellipse with semi-axes
    ``A`` (major) and ``B`` (minor) in mm at angle ``phi_deg`` in physical
    ``(dim0, dim1)`` space. ``s0, s1`` are the mm spacings of dim0, dim1."""
    n0, n1 = shape
    c0, c1 = center_idx
    ii, jj = np.meshgrid(np.arange(n0), np.arange(n1), indexing="ij")
    d0 = (ii - c0) * s0
    d1 = (jj - c1) * s1
    phi = np.deg2rad(phi_deg)
    u = d0 * np.cos(phi) + d1 * np.sin(phi)      # along physical major
    v = -d0 * np.sin(phi) + d1 * np.cos(phi)     # along physical minor
    return ((u / A) ** 2 + (v / B) ** 2 <= 1.0).astype(np.uint8)


def _fit(planner, mask, pixel_sizes, slice_dim=2, slice_idx=0):
    return getattr(planner, _FIT)(
        mask, 1, np.asarray(pixel_sizes, dtype=float), slice_dim, slice_idx
    )


def _physical_axis(landmark, key_a, key_b, voxel):
    """Length (mm) and in-plane direction (deg, mod 180) of the landmark axis,
    mirroring ``_cal_distance`` (``point * voxel_sizes``)."""
    pa = np.asarray(landmark[key_a], dtype=float) * np.asarray(voxel)
    pb = np.asarray(landmark[key_b], dtype=float) * np.asarray(voxel)
    vec = (pa - pb)[:2]
    length = float(np.linalg.norm(pa - pb))
    angle = float(np.degrees(np.arctan2(vec[1], vec[0])) % 180.0)
    return length, angle


def test_anisotropic_recovers_physical_major():
    """Discriminating case: the lesion is longest in *physical* mm along dim0
    (2A = 160 mm), but because dim1 is finer (0.5 mm) it is longest in *pixels*
    along dim1. A correct (real-space) fit must report the dim0 axis as major;
    the pre-fix code either dropped the lesion or labelled the pixel-long axis."""
    s0, s1 = 1.0, 0.5
    A, B, phi = 80.0, 50.0, 25.0          # physical semi-axes (mm), tilt (deg)
    mask = _rasterize_physical_ellipse((220, 320), (110, 160), s0, s1, A, B, phi)
    voxel = (s0, s1, 1.0)                 # slice_dim=2 -> in-plane (dim0, dim1)

    landmarks, _, n_clusters = _fit(_make_planner(), mask, (s0, s1))
    assert n_clusters == 1, f"expected 1 cluster, got {n_clusters}"
    assert len(landmarks) == 1, f"expected 1 landmark, got {len(landmarks)}"

    major_mm, major_dir = _physical_axis(landmarks[0], "P1", "P2", voxel)
    minor_mm, _ = _physical_axis(landmarks[0], "P3", "P4", voxel)

    # Recovered physical lengths match the truth (tolerance covers rasterization
    # + algebraic-conic vs. truth ellipse fit).
    assert abs(major_mm - 2 * A) < 0.08 * 2 * A, f"major {major_mm:.1f} != ~160 mm"
    assert abs(minor_mm - 2 * B) < 0.08 * 2 * B, f"minor {minor_mm:.1f} != ~100 mm"
    # L-1-2 is the physical major (A2): longer than L-3-4.
    assert major_mm > minor_mm, f"major {major_mm:.1f} !> minor {minor_mm:.1f}"
    # Major axis points along the PHYSICAL major direction (~25 deg), NOT the
    # pixel-major direction (~115 deg) the buggy comparison would pick.
    d_phys = min(abs(major_dir - phi), 180 - abs(major_dir - phi))
    assert d_phys < 12.0, f"major_dir {major_dir:.1f} not aligned with {phi} deg"


def test_isotropic_is_a_noop():
    """With isotropic in-plane spacing the fix is a provable no-op: the four
    landmark index coordinates must be byte-identical to the released behaviour.
    Golden captured from the pre-fix code on the same mask."""
    GOLDEN = {
        "P1": [182, 194, 0],
        "P2": [38, 126, 0],
        "P3": [131, 115, 0],
        "P4": [89, 205, 0],
    }
    mask = _rasterize_physical_ellipse((220, 320), (110, 160), 1.0, 1.0, 80.0, 50.0, 25.0)
    landmarks, _, _ = _fit(_make_planner(), mask, (1.0, 1.0))
    assert len(landmarks) == 1, f"expected 1 landmark, got {len(landmarks)}"
    for key, expected in GOLDEN.items():
        assert landmarks[0][key] == expected, (
            f"{key}: {landmarks[0][key]} != golden {expected} (isotropic must be unchanged)"
        )


def test_physical_measurement_is_continuous_and_ordered():
    """The reported L-1-2 / L-3-4 must be the CONTINUOUS real-space ellipse axes,
    persisted in ``landmark['measurements']`` — NOT a re-derivation from the
    rounded landmark points. On near-circular lesions (A ~ B) the rounded points
    can flip order, but the physical measurement must keep major (L-1-2) >= minor
    (L-3-4) and must equal the continuous axis (~2A / ~2B), proving the int
    landmarks are display-only and the rounding never corrupts the measurement."""
    planner = _make_planner()
    # Same well-formed geometry as the other tests, but near-circular (A ~ B) on
    # a few spacings/angles that would stress a rounding tie-break.
    for s0, s1, A, B, phi in [
        (1.0, 0.7, 60.0, 59.0, 30.0),
        (0.8, 1.3, 60.0, 59.2, 70.0),
        (1.0, 1.0, 60.0, 59.0, 15.0),
    ]:
        mask = _rasterize_physical_ellipse((220, 320), (110, 160), s0, s1, A, B, phi)
        landmarks, _, _ = _fit(planner, mask, (s0, s1))
        assert len(landmarks) == 1, f"({s0},{s1},{A},{B},{phi}): got {len(landmarks)} landmarks"
        meas = landmarks[0].get("measurements")
        assert meas is not None, "fit must persist a physical-space 'measurements' field"
        # major >= minor on the PHYSICAL measurement (max/min of axes_real).
        assert meas["L-1-2"] + 1e-9 >= meas["L-3-4"], (
            f"({s0},{s1},{A},{B},{phi}): physical major {meas['L-1-2']:.3f} < minor {meas['L-3-4']:.3f}"
        )
        # The measurement is the continuous axis (~2A / ~2B), within fit tolerance.
        assert abs(meas["L-1-2"] - 2 * A) < 0.08 * 2 * A, f"major {meas['L-1-2']:.2f} != ~{2 * A}"
        assert abs(meas["L-3-4"] - 2 * B) < 0.08 * 2 * B, f"minor {meas['L-3-4']:.2f} != ~{2 * B}"


if __name__ == "__main__":
    test_anisotropic_recovers_physical_major()
    test_isotropic_is_a_noop()
    test_physical_measurement_is_continuous_and_ordered()
    print("OK: ellipse-fit tests passed (anisotropic recovery + isotropic no-op + physical measurement)")