Lacunae
Accelerated MRI Reconstruction from Undersampled K-Space via U-Net
Highlights
- ❀ Systematically compares 4 k-space sampling patterns across 3 acceleration factors on a fixed U-Net, isolating sampling strategy as an independent variable
- ❀ Trained and evaluated on the fastMRI single-coil knee dataset with SSIM and PSNR as primary metrics
- ❀ 12 total training runs spanning the full pattern × acceleration grid
- ❀ Addresses a gap in the literature, where random Cartesian undersampling is used by default without direct comparison to alternatives
Most accelerated MRI reconstruction papers default to random Cartesian undersampling without justification. This study asks a question that hasn't been answered cleanly for deep-learning-based reconstruction: does the choice of k-space sampling pattern matter as much as — or more than — model architecture?
We train the same U-Net across four mask types and three acceleration factors, producing a systematic comparison of reconstruction quality that the literature currently lacks.
Study Design
| Variable | Values |
|---|---|
| Sampling patterns | Random Cartesian, Equispaced, Radial, Variable-density |
| Acceleration factors | 4x, 8x, 16x |
| Model | U-Net (fixed architecture across all runs) |
| Metrics | SSIM, PSNR |
| Dataset | fastMRI single-coil knee |
12 total training runs. Results reported as mean SSIM and PSNR on the held-out test split.
Sampling Patterns
- Random Cartesian — the de facto standard. Randomly samples k-space columns while always retaining the center low-frequency region. Used as the baseline for comparison.
- Equispaced — samples every Nth column uniformly. Deterministic and hardware-friendly, but produces coherent aliasing artifacts.
- Radial — samples columns with probability weighted by distance from the k-space center, simulating radial acquisition. Denser at low frequencies, where image energy is concentrated.
- Variable-density — quadratic falloff from center, providing a smooth transition between dense low-frequency sampling and sparse high-frequency sampling.
Model
A standard U-Net with four encoder/decoder stages, held fixed across all experimental conditions so that observed differences in reconstruction quality are attributable to the sampling pattern, not the model.
- Encoder: ConvBlock(1→32) → ConvBlock(32→64) → ConvBlock(64→128) → ConvBlock(128→256)
- Bottleneck: ConvBlock(256→512)
- Decoder: mirrors encoder with transposed convolutions and skip connections
- Loss: L1 + SSIM
- Parameters: ~7.7M
Status
Random Cartesian baseline trained and evaluated at 4x and 8x acceleration. Equispaced, radial, and variable-density mask implementations, the full 4×3 experimental grid, and the final results table and analysis are still in progress.
Stack