Notes on SURE

[AnderBiguri]

• SURE assumes Gaussian noise, but FDK is not. SURE formula for noise in the exponential family?
• SURE in its basic form needs an identity linear model. Generalized SURE is the one for other models
• Generalized SURE requires P, which is not trivial to obtain for CT. However Tatianna Bubba paper suggest that for limited angle CT, it could be computable.

sources:
Unsupervised Learning with Stein’s Unbiased Risk Estimator
Deep neural networks for inverse problems with pseudodifferential operators: An application to limited-angle tomography

[AnderBiguri]

More notes: deepinv tackles most of those things above.
Read more at:
https://deepinv.github.io/deepinv/deepinv.loss.html#self-supervised-learning

sources:
https://ieeexplore.ieee.org/abstract/document/6714502

For uknown noise levels: UNSURE:
https://arxiv.org/abs/2409.01985