Table 7 The results and their standard errors for multi-task. The first block shows the results for corrupting the original images by subsampling the k-space and applying motion artefacts. The second block shows the results for correcting images that are corrupted by noise and bias fields as well. The results in bold indicate the best performance without significant differences between them following a Nemenyi post-hoc test
From: Improving MR image quality with a multi-task model, using convolutional losses
SSIM | VIF | |
|---|---|---|
Original | \(0.625\pm 0.162\) | \(0.905\pm 0.075\) |
Subsampling | \(0.662\pm 0.138\) | \(0.953\pm 0.076\) |
Subsampling + Motion | \(0.633\pm 0.157\) | \(0.921\pm 0.082\) |
Motion | \(0.678\pm 0.129\) | \(0.919\pm 0.076\) |
Motion + Subsampling | \(0.631\pm 0.159\) | \(\varvec{0.963\pm 0.074}\) |
MT | \(0.672\pm 0.148\) | \(0.930\pm 0.107\) |
MT+\(\mathcal {L}_C\) | \(\varvec{0.692\pm 0.137}\) | \(0.941\pm 0.080\) |
Original | \(0.380\pm 0.080\) | \(0.845\pm 0.070\) |
Bias | \(0.399\pm 0.072\) | \(0.810\pm 0.075\) |
Bias + Noise | \(0.764\pm 0.137\) | \(\varvec{1.010\pm 0.049}\) |
Noise | \(0.788\pm 0.113\) | \(1.009\pm 0.049\) |
Noise + Bias | \(0.420\pm 0.072\) | \(0.832\pm 0.065\) |
MT | \(0.792\pm 0.106\) | \(0.986\pm 0.048\) |
MT+\(\mathcal {L}_C\) | \(\varvec{0.796\pm 0.115}\) | \(\varvec{1.007\pm 0.049}\) |