Efficient Noise Calculation in Deep Learning-based MRI Reconstructions
We introduce a practical technique to measure voxel-wise noise variance in AI-based MRI reconstructions, addressing a long-standing gap in accelerated MRI. Traditional noise analysis—essential for judging reconstruction fidelity and diagnostic reliability—has been largely neglected in deep learning methods due to the analytical and computational burden of modeling noise propagation.
Our method uses a Jacobian sketching estimator to efficiently approximate voxel-level variance, achieving accuracy comparable to Monte-Carlo simulations while requiring far less computation and memory. Demonstrated on knee and brain MRI across various models, undersampling patterns, and noise levels, this approach makes precise noise quantification feasible again, supporting more reliable deployment of accelerated AI reconstructions in clinical practice.
Dalmaz, O., Desai, A., Heckel, R., Çukur, T., Chaudhari, A., & Hargreaves, B. (2025). Efficient Noise Calculation in Deep Learning-based MRI Reconstructions. ICML 2025.
Example of accelerated MRI reconstruction with voxel-wise noise variance estimation. Our method efficiently produces noise maps (heatmap overlays) that match Monte-Carlo references at a fraction of the computational and memory cost, enabling reliable assessment of image fidelity in deep learning–based MRI