MRI Reconstruction Losses#

class atommic.collections.reconstruction.losses.NoiseAwareLoss(*args: Any, **kwargs: Any)[source]#

Bases: Loss

Computes the Noise Aware loss between two tensors.

Note

Extends atommic.core.classes.loss.Loss.

Examples

>>> from atommic.collections.reconstruction.losses.na import NoiseAwareLoss
>>> import torch
>>> loss = NoiseAwareLoss(win_size=7, k1=0.01, k2=0.03)
>>> loss(X=torch.rand(1, 1, 256, 256), Y=torch.rand(1, 1, 256, 256))
tensor(0.0872)
forward(target: torch.Tensor, pred: torch.Tensor, mask: Optional[torch.Tensor] = None, sigma: float = 0.0) → torch.Tensor[source]#

Forward pass of NoiseAwareLoss.

Parameters

  • target (torch.Tensor) – The target tensor.

  • pred (torch.Tensor) – The predicted tensor.

  • mask (torch.Tensor) – The mask tensor. If None, all pixels are considered.

  • sigma (float) – The noise level.

class atommic.collections.reconstruction.losses.SSIMLoss(*args: Any, **kwargs: Any)[source]#

Bases: Loss

Computes the (1-) SSIM loss between two tensors.

Examples

>>> from atommic.collections.reconstruction.losses.ssim import SSIMLoss
>>> import torch
>>> loss = SSIMLoss(win_size=7, k1=0.01, k2=0.03)
>>> loss(X=torch.rand(1, 1, 256, 256), Y=torch.rand(1, 1, 256, 256), data_range=torch.tensor([1.]))
tensor(0.9872)
__init__(win_size: int = 7, k1: float = 0.01, k2: float = 0.03)[source]#

Inits SSIMLoss.

Parameters

  • win_size (int, optional) – Window size for SSIM calculation.

  • k1 (float, optional) – k1 parameter for SSIM calculation.

  • k2 (float, optional) – k2 parameter for SSIM calculation.

forward(X: torch.Tensor, Y: torch.Tensor, data_range: Optional[torch.Tensor] = None)[source]#

Forward pass of SSIMLoss.

Parameters

  • X (torch.Tensor) – First input tensor.

  • Y (torch.Tensor) – Second input tensor.

  • data_range (torch.Tensor) – Data range of the input tensors. If None, it is computed as the maximum range of the input tensors. Default is None.