Flow-based generative model slashes MRI scan data by 8x with high accuracy
New AI method achieves 29.24 dB PSNR on 8x accelerated MRI using only 5% of measurements
Deep Dive
This paper presents a task-aware flow-based generative framework that learns optimal subsampling masks for compressed sensing. The model achieves state-of-the-art image reconstruction with a PSNR of 25.17 dB at 5% subsampling on CelebA and 29.24 dB on 8× accelerated MRI from fastMRI. It reformulates conventional Flow Matching training to enhance performance in image classification, reconstruction, and MRI acceleration with minimal computational overhead.
Key Points
- Achieves PSNR of 25.17 dB at 5% subsampling rate on CelebA image reconstruction
- Reaches 29.24 dB PSNR on 8x accelerated MRI scans (fastMRI dataset) with minimal overhead
- Reformulates Flow Matching training to learn task-aware subsampling masks for classification, reconstruction, and acceleration
Why It Matters
Faster, higher-quality medical imaging with fewer measurements could reduce scan times and costs in clinical settings.