Novel Algorithm for Real-Time MRI Reconstruction
Introduction
Deep learning methods accelerate MRI reconstruction but often require large training datasets. We propose a lightweight convolutional network that achieves <10ms reconstruction time with only 50 training pairs.
Methods
We trained a 3-layer CNN on undersampled k-space data from 50 subjects. Data augmentation was used to prevent overfitting. Reconstruction was performed on a single GPU.
Results
Our method achieved PSNR of 38.2 dB and SSIM of 0.94, outperforming compressed sensing by 12% in reconstruction time. Visual inspection showed preserved fine details.
Conclusion
Our approach enables real-time MRI reconstruction with minimal training data, making it suitable for clinical deployment. Future work will explore multi-coil extensions.
Ask Atty to draft your research poster now.