Class GibbsNoise

monai/transforms/intensity/array.py:1926–1999

The transform applies Gibbs noise to 2D/3D MRI images. Gibbs artifacts are one of the common type of type artifacts appearing in MRI scans. The transform is applied to all the channels in the data. For general information on Gibbs artifacts, please refer to: `An Image-based A

class GibbsNoise(Transform, Fourier):
    """
    The transform applies Gibbs noise to 2D/3D MRI images. Gibbs artifacts
    are one of the common type of type artifacts appearing in MRI scans.

    The transform is applied to all the channels in the data.

    For general information on Gibbs artifacts, please refer to:

    `An Image-based Approach to Understanding the Physics of MR Artifacts
    <https://pubs.rsna.org/doi/full/10.1148/rg.313105115>`_.

    `The AAPM/RSNA Physics Tutorial for Residents
    <https://pubs.rsna.org/doi/full/10.1148/radiographics.22.4.g02jl14949>`_

    Args:
        alpha: Parametrizes the intensity of the Gibbs noise filter applied. Takes
            values in the interval [0,1] with alpha = 0 acting as the identity mapping.
    """

    backend = [TransformBackends.TORCH, TransformBackends.NUMPY]

    def __init__(self, alpha: float = 0.1) -> None:
        if alpha > 1 or alpha < 0:
            raise ValueError("alpha must take values in the interval [0, 1].")
        self.alpha = alpha

    def __call__(self, img: NdarrayOrTensor) -> NdarrayOrTensor:
        img = convert_to_tensor(img, track_meta=get_track_meta())
        img_t = convert_to_tensor(img, track_meta=False)
        n_dims = len(img_t.shape[1:])

        # FT
        k = self.shift_fourier(img_t, n_dims)
        # build and apply mask
        k = self._apply_mask(k)
        # map back
        out = self.inv_shift_fourier(k, n_dims)
        img, *_ = convert_to_dst_type(out, dst=img, dtype=out.dtype)

        return img

    def _apply_mask(self, k: NdarrayOrTensor) -> NdarrayOrTensor:
        """Builds and applies a mask on the spatial dimensions.

        Args:
            k: k-space version of the image.
        Returns:
            masked version of the k-space image.
        """
        shape = k.shape[1:]

        # compute masking radius and center
        r = (1 - self.alpha) * np.max(shape) * np.sqrt(2) / 2.0
        center = (np.array(shape) - 1) / 2

        # gives list w/ len==self.dim. Each dim gives coordinate in that dimension
        coords = np.ogrid[tuple(slice(0, i) for i in shape)]