Efficient Noise Calculation in Deep Learning-based MRI Reconstructions
Authors: Onat Dalmaz, Arjun D Desai, Reinhard Heckel, Tolga Cukur, Akshay S Chaudhari, Brian Hargreaves
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method on knee and brain MRI datasets for both data- and physics-driven networks trained in supervised and unsupervised manners. Compared to empirical references obtained via Monte-Carlo simulations, our technique achieves near-equivalent performance while reducing computational and memory demands by an order of magnitude or more. Furthermore, our method is robust across varying input noise levels, acceleration factors, and diverse undersampling schemes, highlighting its broad applicability. Our work reintroduces accurate and efficient noise analysis as a central tenet of reconstruction algorithms, holding promise to reshape how we evaluate and deploy DL-based MRI. |
| Researcher Affiliation | Academia | 1Department of Electrical Engineering, Stanford University, Stanford, CA, USA 2Department of Radiology, Stanford University, Stanford, CA, USA 3Department of Computer Engineering, Technical University of Munich, Munich, Germany 4Department of Electrical and Electronics Engineering, Bilkent University, Ankara, Turkey 5Department of Biomedical Data Science, Stanford University, Stanford, CA, USA 6Department of Bioengineering, Stanford University, Stanford, CA, USA. Correspondence to: Onat Dalmaz <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 Naive Per-Voxel Variance Calculation (Row-by-Row Jacobian) Algorithm 2 Noise Calculation in DL-based MRI reconstruction via Sketching the Network Jacobian |
| Open Source Code | Yes | Our code is available at https://github.com/onatdalmaz/deep recon noise. |
| Open Datasets | Yes | We performed experiments on two publicly available datasets Stanford knee dataset and a subset of the fast MRI brain dataset (Knoll et al., 2020c). ... fastmri: A publicly available raw k-space and dicom dataset of knee images for accelerated mr image reconstruction using machine learning. Radiology: Artificial Intelligence, 2(1):e190007, 2020c. |
| Dataset Splits | Yes | The Stanford knee dataset consists of 8-channel 3D FSE PD-weighted scans, which we split into 14 subjects for training, 2 for validation, and 3 for testing. Each 3D scan was demodulated and decoded via a 1D inverse Fourier transform along the readout dimension, yielding 2D Axial slices with matrix size 320 256. The fast MRI brain dataset comprises 16-channel Axial T2-weighted scans with matrix size 384 384, split into 54 training, 20 validation, and 30 testing subjects. |
| Hardware Specification | Yes | We used Py Torch for modeling (Paszke et al., 2019), and NVIDIA NVIDIA RTX A6000 for GPU acceleration. |
| Software Dependencies | No | We used Py Torch for modeling (Paszke et al., 2019), and NVIDIA NVIDIA RTX A6000 for GPU acceleration. The paper mentions PyTorch but does not specify a version number for it or any other software dependency. |
| Experiment Setup | Yes | We configure E2E-Var Net with 4 unrolled steps, each containing 2 Res Net blocks (256 channels, kernel size 3). An ℓ1-norm reconstruction loss is used, and the network is optimized with Adam (lr=10 4). ... Tables 3, 4, 5, 6, 7, and 8 provide detailed hyperparameters for each method, including unrolled steps, block architecture, number of ResNet blocks per step, kernel size, loss function, batch size, base learning rate, weight decay, and maximum iterations. |