[10/24/23] Aim 3: Posterior Sampling and Uncertainty

y \sim \text{Poission}(I_0~e^{-Ax})

y \sim \text{Poission}(I_0~e^{-Ax})

To compare different sampling methods in terms of uncertainty as a function of:

Dose (i.e. $I_0$ ) $\quad\quad$ and $\quad\quad$ Number of views

Baseline methods:

1. FBP

2. FBP + U-net with quantile regression (in progress)

Non-linear DPS + {Poisson, Gaussian} Measurement:

- Fixed DPS step @ 0.3

- 20 validation images

- 120 samples per image to compute uncertainty

- 500 steps ( $\approx$ 3 mins / 12 parallel samples)

Trying to reduce to 200 steps with DDIM

FBP

poisson

gaussian

$I_0 = 4096$

$I_0 = 512$

$I_0 = 128$

FBP

poisson

gaussian

$I_0 = 4096$

$I_0 = 512$

$I_0 = 128$