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?