AI in MR Image Formation

The core MRI problem (3-way trade-off):

• Scan time ↔ Resolution ↔ SNR (Signal-to-Noise Ratio)
• Long scan times → patient discomfort, motion artefacts, high cost

MRI data acquisition — k-space:

• Raw MRI data is acquired in Fourier (k-) space, not image space
• Fully sampled k-space → clean image; undersampled k-space → artefacts
• Scan acceleration = collecting fewer k-space lines (e.g., R=4 means 4× fewer measurements)

3 approaches to AI-based MRI reconstruction:

1. Image processing based — treat reconstruction as a denoising/post-processing problem

2. k-space recovery — directly recover missing data in the frequency domain

3. DL iterative reconstruction — unroll physics-based gradient descent steps into a neural network (learns T gradient descent steps end-to-end)

DL iterative reconstruction (most important approach):

• Physics (known MRI forward model) + Learning combined
• "Unrolled iterations mapped on neural network"
• Key paper: Hammernik MRM 2018
• Result: Reconstruction network outperforms image processing (denoising) network in SSIM

fastMRI dataset (open benchmark):

• Created by NYU + Facebook Research
• Contains: Knee (1398 cases), Brain (7002 cases), Prostate (312 cases), Breast (300 cases)
• ~9,000 visitors/year, 961 TB downloaded/year

Key challenges & limitations of AI MRI reconstruction:

• Hallucinations — AI can generate anatomical structures that aren't there (dangerous in clinical setting)
• Out-of-domain generalization — model trained on one scanner may fail on another
• Bayesian Uncertainty Estimation — used to flag uncertain reconstructions