Enhancement of arterial spin labelling MR images using deep learning.
Current plan: CNN trained to map from (low-repetition, noisy and artifact-prone ASL images + M0 image + T1 image) to (high-quality, filtered/smoothed perfusion map).
Architecture ideas:
- CNN
- GAN?
- Treat time (fourth dimension) as a fourth spatial dimension.
Loss function ideas:
- Voxelwise loss
- Adversarial loss?
- Perceptual loss?