RecoveryCoefficient#

class deepinv.loss.metric.RecoveryCoefficient(eps=None, **kwargs)[source]#

Bases: Metric

Recovery Coefficient metric used in emission tomography.

Computes the ratio between the total reconstructed activity and the total ground-truth activity inside a given region of interest, defined by a mask:

\[RC = \frac{\sum_i \hat{x}_i m_i} {\sum_i x_i m_i + \varepsilon}\]

where \(\hat{x}\) is the reconstructed image, \(x\) is the ground-truth image, \(m\) is a binary or weighted mask defining the region of interest, and \(\varepsilon\) is a small constant added for numerical stability.

A value of 1 indicates perfect recovery of activity within the masked region. Values below 1 indicate underestimation, while values above 1 indicate overestimation.

Note

This metric requires a mask keyword argument to be provided during evaluation.

Note

Higher values are better when they are closer to 1. Therefore, lower_better=False.

Example:

>>> import torch
>>> from deepinv.loss.metric import RecoveryCoefficient
>>> m = RecoveryCoefficient()
>>> x = torch.ones(2, 1, 4, 4)
>>> x_net = torch.ones(2, 1, 4, 4)
>>> mask = torch.ones_like(x)
>>> m(x_net, x, mask=mask)
tensor([1., 1.])
Parameters:

  • eps (float) – Small constant added to the denominator for numerical stability.

  • kwargs – Additional keyword arguments passed to the parent deepinv.loss.metric.Metric class.

invert_metric(m)[source]#

“Invert metric for use as a training loss. Recovery Coefficient is optimal at 1 so a sign flip does not produce a valid loss :param torch.Tensor m: calculated metric