Gaussian surrogates do well on Poisson inverse problems
New research finds simpler Gaussian surrogates achieve comparable MSE to complex Poisson models in medical CT scans.
Researchers Alexandra Spitzer, Lorenzo Baldassari, Valentin Derbanot, and Ivan Dokmanić published a paper showing that Gaussian surrogate objectives perform nearly as well as Poisson maximum a posteriori (MAP) estimators for low-dose imaging problems. In computed tomography with Poisson noise, their simpler linear estimators achieved mean squared error (MSE) comparable to more complex Poisson models. This suggests practitioners could use faster, simpler algorithms without sacrificing reconstruction quality in medical and scientific imaging.
Why It Matters
Enables faster, simpler image reconstruction for low-dose medical scans like CT, reducing computational costs without losing accuracy.