https://atlas.rsna.org/schemas/2025-11/model.json

Zero-filled Skipped Connection Network (ZSNET) for 2D Brain MRI Reconstruction

https://pubs.rsna.org/doi/10.1148/ryai.210313

1.0

moc.atem@yelkcumm

National Institutes of Health (R01EB024532, P41EB017183).

Retrospective study with local IRB approval; informed consent waived due to minimal risk and use of de-identified data.

Enlarged End-to-End Variational Network with 18 cascades; each cascade uses a residual U-Net (20 channels, five pooling layers) and a DenseNet-inspired skipped connection to the zero-filled reconstruction; coil sensitivities estimated using all k-space. Approx. 231M parameters. Single-direction phase-encode undersampling.

Open-source code available at https://github.com/facebookresearch/fastMRI

Substantially accelerates 2D brain MRI acquisition while preserving diagnostic image quality up to approximately 4× and enabling potential screening use up to approximately 14× in many cases, reducing scan times and improving patient comfort and throughput.

Clinical imaging acquisition/reconstruction and workflow optimization.

Automatic image reconstruction from undersampled multi-coil k-space without user intervention.

Reconstruction of undersampled axial 2D fast spin-echo T1-weighted, T2-weighted, and T2 FLAIR brain MRI acquired on 1.5T and 3T Siemens scanners; general diagnostic imaging up to ~4× acceleration and potential screening up to ~14× based on radiologist evaluation.

Raw multi-coil k-space from 2D spin-echo brain MRI undersampled along a single phase-encoding direction (accelerations trained at 2×–100×).

Model trained for accelerations 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 24, 32, 48, 64, 100× using fastMRI brain training/validation splits; evaluate on held-out test set. For reader studies, low-level dithering noise was added before presentation.

Single-institution, single-vendor dataset (Siemens); limited sequences (2D FSE T1, T2, T2 FLAIR) and section thickness (3–5 mm); retrospective undersampling; quantitative metrics (SSIM/PSNR) do not fully distinguish screening vs diagnostic thresholds; pseudonormalization, blurring, and potential hallucinations at high accelerations; out-of-distribution performance degrades (e.g., knees).

For general-purpose diagnostic imaging use up to approximately 4× acceleration; for potential screening, many cases acceptable up to approximately 14×; higher accelerations may risk pseudonormalization and loss of subtle findings.

Uses publicly available fastMRI dataset with standard train/validation/test splits; code available; evaluation with bootstrap CIs and detailed supplemental methods.