Table 2 Results of the Bayesian hyper-parameter optimization. The values marked with “\(^*\)” were not optimized, since the \(\delta\) parameters are only used in the models that implement the convolutional loss (\(\mathcal {L}_C\))
From: Improving MR image quality with a multi-task model, using convolutional losses
Hyper-p. | Models | ||||||
|---|---|---|---|---|---|---|---|
Bias | Subsampling | Motion | Noise | MT | MT+\(\mathcal {L}_C\) | Pelvis MT+\(\mathcal {L}_C\) | |
Optim. | RMSprop | RMSprop | RMSprop | RMSprop | RMSprop | RMSprop | RMSprop |
L.r., \(10^{\gamma }\) | \(-3.79\) | \(-4.39\) | \(-4.17\) | \(-4.06\) | \(-3.73\) | \(-3.91\) | \(-3.88\) |
\(\alpha\) | 0.60 | 0.48 | 0.59 | 0.85 | 0.51 | 0.87 | 0.89 |
\(\delta _{E}\) | \(1^*\) | \(1^*\) | \(1^*\) | \(1^*\) | \(1^*\) | 0.30 | 0.13 |
\(\delta _{PT}\) | \(0^*\) | \(0^*\) | \(0^*\) | \(0^*\) | \(0^*\) | 0.83 | 0.23 |
\(\delta _{PR}\) | \(0^*\) | \(0^*\) | \(0^*\) | \(0^*\) | \(0^*\) | 0.25 | 0.03 |
\(\delta _{S3T}\) | \(0^*\) | \(0^*\) | \(0^*\) | \(0^*\) | \(0^*\) | 0.74 | 0.84 |
\(\delta _{S3R}\) | \(0^*\) | \(0^*\) | \(0^*\) | \(0^*\) | \(0^*\) | 0.30 | 0.23 |
\(\delta _{S5T}\) | \(0^*\) | \(0^*\) | \(0^*\) | \(0^*\) | \(0^*\) | 0.21 | 0.14 |
\(\delta _{S5R}\) | \(0^*\) | \(0^*\) | \(0^*\) | \(0^*\) | \(0^*\) | 0.72 | 0.91 |
\(\delta _{L3}\) | \(0^*\) | \(0^*\) | \(0^*\) | \(0^*\) | \(0^*\) | 0.09 | 0.38 |
\(\delta _{L5}\) | \(0^*\) | \(0^*\) | \(0^*\) | \(0^*\) | \(0^*\) | 0.82 | 0.77 |