Problem
Accelerated MRI reconstructs an image from undersampled k-space. Classical compressed sensing works; unrolled networks try to keep the optimization structure and learn the parts that used to be hand-designed.
Approach
I implemented ISTA-Net and ADMM-Net for reconstruction from 25% undersampled k-space, then extended ADMM-Net with Squeeze-and-Excitation blocks.
- Unrolled optimization with learned proximal operators instead of a generic image-to-image CNN.
- SE-block extension improved convergence speed (1.14×) without dropping reconstruction quality.
- Full pipeline: k-space loading, training, PSNR/SSIM evaluation, and ablation studies, landing at 24 dB PSNR on the evaluation set.
The point of unrolling is interpretability you can still train. The SE block was a small inductive-bias bet that paid off on wall-clock, not on a prettier loss curve.