Class RandBiasField

monai/transforms/intensity/array.py:746–836

Random bias field augmentation for MR images. The bias field is considered as a linear combination of smoothly varying basis (polynomial) functions, as described in `Automated Model-Based Tissue Classification of MR Images of the Brain <https://ieeexplore.ieee.org/stamp/stamp.jsp?tp

class RandBiasField(RandomizableTransform):
    """
    Random bias field augmentation for MR images.
    The bias field is considered as a linear combination of smoothly varying basis (polynomial)
    functions, as described in `Automated Model-Based Tissue Classification of MR Images of the Brain
    <https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=811270>`_.
    This implementation adapted from `NiftyNet
    <https://github.com/NifTK/NiftyNet>`_.
    Referred to `Longitudinal segmentation of age-related white matter hyperintensities
    <https://www.sciencedirect.com/science/article/pii/S1361841517300257?via%3Dihub>`_.

    Args:
        degree: degree of freedom of the polynomials. The value should be no less than 1.
            Defaults to 3.
        coeff_range: range of the random coefficients. Defaults to (0.0, 0.1).
        dtype: output data type, if None, same as input image. defaults to float32.
        prob: probability to do random bias field.

    """

    backend = [TransformBackends.NUMPY]

    def __init__(
        self,
        degree: int = 3,
        coeff_range: tuple[float, float] = (0.0, 0.1),
        dtype: DtypeLike = np.float32,
        prob: float = 0.1,
    ) -> None:
        RandomizableTransform.__init__(self, prob)
        if degree < 1:
            raise ValueError(f"degree should be no less than 1, got {degree}.")
        self.degree = degree
        self.coeff_range = coeff_range
        self.dtype = dtype

        self._coeff = [1.0]

    def _generate_random_field(self, spatial_shape: Sequence[int], degree: int, coeff: Sequence[float]):
        """
        products of polynomials as bias field estimations
        """
        rank = len(spatial_shape)
        coeff_mat = np.zeros((degree + 1,) * rank)
        coords = [np.linspace(-1.0, 1.0, dim, dtype=np.float32) for dim in spatial_shape]
        if rank == 2:
            coeff_mat[np.tril_indices(degree + 1)] = coeff
            return np.polynomial.legendre.leggrid2d(coords[0], coords[1], coeff_mat)
        if rank == 3:
            pts: list[list[int]] = [[0, 0, 0]]
            for i in range(degree + 1):
                for j in range(degree + 1 - i):
                    for k in range(degree + 1 - i - j):
                        pts.append([i, j, k])
            if len(pts) > 1:
                pts = pts[1:]
            np_pts = np.stack(pts)
            coeff_mat[np_pts[:, 0], np_pts[:, 1], np_pts[:, 2]] = coeff