Fast MRI scanners save time by measuring only a fraction of the raw frequency-domain data, called k-space, and then letting a neural network fill in the rest. The best current systems learn which measurements to take jointly with the network that reconstructs the image. They all assume that every measurement is equally noisy. Real acquisitions break this assumption in structured, scan-specific ways. Radio-frequency interference can contaminate particular frequency bands, coil noise levels change with patient loading, and motion can corrupt particular readout lines. Statistics has a clear answer for this situation. When noise varies across possible measurements, you should spend your budget where noise is low. Active Heteroscedastic Regression proves that preferring such "quiet" samples improves the error rate by a factor of the dimension. Learned MRI acquisition has never used this, because its noise model has no way to express it.
There is an obstacle. Doing better here requires the sampling pattern to change from scan to scan. Bakker et al. found that when learned sampling patterns adapt to each image's content, reconstruction gets worse, and their best policy collapsed to a single fixed mask. Their explanation, which they call an "amortization trap", is that a single reconstruction network cannot learn to undo every possible aliasing pattern. This note argues that adapting to noise is different from adapting to anatomy. Noise structure can be estimated before acquisition from a short noise-only prescan plus the always-acquired central k-space lines. It can be summarized in a few numbers and fed to both the sampler and the reconstructor. That gives a family of masks that is low-dimensional and labeled by those numbers, rather than an unstructured, content-driven family. We propose to build and test exactly this, and to find out when it beats much simpler fixes.
Core Claim
Our claim is that a joint sampler and reconstructor, both conditioned on an explicit per-scan noise-variance descriptor, can turn heteroscedastic noise into a reconstruction gain without falling into the amortization trap. Heteroscedastic noise means noise whose variance differs from one measurement to another.
The gain should have three properties. It should grow with how unequal the noise variances are. It should vanish when the noise is uniform, recovering fixed learned masks such as LOUPE. And it should come mainly from the sampler choosing different k-space lines, not only from the reconstructor down-weighting noisy ones.
Cartesian MRI offers a concrete route to the "quiet advantage". In Cartesian sampling, the scanner acquires full horizontal lines of k-space. MR images are nearly real-valued, so k-space is approximately conjugate-symmetric. A noisy line at frequency $+k$ can therefore often be replaced by its mirror at $-k$ at almost no cost in information. More generally, the sampler can trade a noisy line for a nearby quiet one. Fixed masks cannot make this trade, because they do not know which lines are noisy on a given scan.
The novelty is specific. Prior joint acquisition work, including LOUPE, Weiss et al., DPS, BJORK, sequential sampling and Bakker et al.'s policies, learns masks under a uniform-noise fidelity term. Those systems adapt, if at all, to image content. The closest heteroscedastic sampling work, Chaudhuri et al., uses a linear variance model, Gaussian covariates and a weighted least-squares fit, with no imaging operator. MASC avoids metal-corrupted k-space using reinforcement learning, but treats the corruption as a deterministic artifact rather than as a noise variance to be estimated and conditioned on.
Research Program
The measurement model is $y = M_S F x + e$. Here $x$ is the image, $F$ is the coil-weighted Fourier operator, $M_S$ selects the acquired lines $S$, and $e$ is Gaussian with a diagonal covariance $\Sigma_\nu$. That covariance is set by a scan-level noise descriptor $\nu$: a small vector giving per-line or per-band variances and per-coil levels, estimated from the prescan and the central lines.
The reconstructor is an unrolled variational network in the style of E2E VarNet. Its data-consistency step weights each measurement by its precision:
$$\hat{x}^{t+1} = \hat{x}^{t} - \lambda_t F^H M_S^\top \Sigma_\nu^{-1}\left(M_S F \hat{x}^{t} - y\right) + R_t(\hat{x}^{t}, \nu).$$
In this update, $\hat{x}^t$ is the image estimate after $t$ steps, $\lambda_t$ is a learned step size, $F^H$ is the adjoint of the Fourier operator, $M_S^\top$ places the acquired lines back into full k-space, and $R_t$ is a learned image-domain regularizer that also receives $\nu$.
Following Immer et al., any learned refinement of the variance is parameterized through precision, which is the natural parameter of the Gaussian, rather than through the standard deviation. Immer et al. showed this choice removes bad local optima in heteroscedastic likelihood fitting. We borrow the technique; we do not claim to inherit their concavity guarantee.
The sampler maps $\nu$, together with the central-line reconstruction, to line logits. A straight-through top-$q$ selection, as in Weiss et al., enforces the acceleration factor exactly. Sampler and reconstructor are trained end-to-end on a supervised image loss.
We will also derive a linear-Gaussian reference. Under a Gaussian image prior with covariance $C$, the expected posterior error of a line set is
$$\mathrm{MSE}(S;\nu) = \mathrm{tr}\left[\left(C^{-1} + F^H M_S^\top \Sigma_\nu^{-1} M_S F\right)^{-1}\right].$$
This expression has two uses. First, greedy minimization of it gives a classical, network-free A-optimal design that serves as a baseline. Second, it lets us extend Chaudhuri et al.'s quiet-advantage argument from regression to a subsampled Fourier operator. We will ask how much expected error can drop, relative to the best noise-blind mask, as a function of the spread of $\Sigma_\nu$ and the degree of conjugate symmetry. This analysis supplies the predicted shape of the empirical curves.
fastMRI does not, to our knowledge, provide ground-truth heteroscedastic noise labels, so heteroscedasticity is simulated on raw fastMRI multi-coil knee data. We use three families of noise profiles. The first is narrowband interference that raises the variance of a random contiguous band of lines. The second is random per-coil noise scaling. The third is per-line variance inflation mimicking motion-corrupted readouts. A single parameter controls the spread of variances in each family.
Baselines
The naive reference is a fixed equispaced mask with a VarNet trained using ordinary uniform-noise data consistency, on the same heteroscedastic training data. Bakker et al. report 91.30 SSIM for equispaced sampling at 8x on clean 4-coil 128×128 knee data. We will first reproduce this setting without added heteroscedasticity, to confirm our pipeline matches theirs.
The strongest existing learned methods are LOUPE's fixed learned mask and Bakker et al.'s adaptive policy network. Bakker et al. report 91.37 SSIM for LOUPE and 93.26 SSIM for their policy at 8x on coil-compressed fastMRI knee data. These are the strongest numbers in the file for this exact setup. Both baselines will be retrained on our heteroscedastic data with the same reconstructor, data split, acceleration, number of gradient steps and hyperparameter search budget.
We will also give both of them precision-weighted data consistency and access to $\nu$ in the reconstructor. This makes the comparison isolate the effect of noise-aware sampling rather than noise-aware reconstruction.
The principled non-learned competitor is the greedy A-optimal design from the expression above. It is computed per scan from $\nu$ and paired with the same noise-conditioned VarNet.
The key ablation is a noise-conditioned reconstructor with a single fixed learned mask. It isolates whether adapting the sampling to noise adds anything beyond weighting measurements correctly.
A second ablation is a bank of LOUPE masks, one per discretized noise profile. It tests whether continuous conditioning matters.
To count as beating these baselines, the proposed system must exceed the strongest of them in SSIM and PSNR at matched acceleration. The margin must clear seed-to-seed variation across at least three seeds. It must also grow monotonically with the variance-spread parameter. A gain only at extreme spreads would be reported as a narrow result.
Why not just…?
Why not just prewhiten and weight the data-consistency step, keeping a fixed mask? This is the most serious threat, and the fixed-mask ablation tests it directly. Weighting by inverse variance makes the best use of whatever was measured. It cannot recover information from a line that was spent on a noise-dominated measurement when a quiet substitute existed. The linear-Gaussian expression lets us predict, before training, how large the gap should be for each noise family. If the fixed-mask ablation matches the full method, we will report that heteroscedastic MRI needs only noise-aware reconstruction. That would be a useful negative result.
Why not just train one mask per noise regime? A mask bank sidesteps the amortization trap entirely, and for coarse, discrete noise conditions it may be as good. The method must win on noise profiles that fall between the bank's training regimes and on unseen interference positions. If the bank matches it there too, the contribution shrinks to "condition the reconstructor". We would report that outcome honestly.
Why not just use classical Bayesian experimental design? A-optimal design given $\nu$ is principled and needs no sampler training. Its weakness is the Gaussian image prior, which ignores the structure that learned reconstructors exploit. The same weakness appears in DPS at 80x, where learned sampling reached 36.2 dB PSNR versus 27.7 dB for random sampling. If greedy design paired with the learned reconstructor matches end-to-end sampling, we will recommend it as the cheaper method.
Why not just apply a variance-stabilizing transform? Transforms such as Anscombe or log scaling are nonlinear. Applied to complex k-space, they break the linear forward model that data consistency relies on. Landa and Kluger's Dyson equalizer also found such transforms insufficient for structured heteroscedastic matrices. We therefore do not run this as a main baseline.
Proposed Experiments
All runs use fastMRI multi-coil knee data, compressed to 4 virtual coils and cropped to 128×128 as in Bakker et al., at 4x and 8x acceleration. The data use the three simulated noise families with the variance-spread parameter swept from zero upward. Metrics are SSIM, PSNR and normalized mean squared error, plus error measured inside the frequency bands the noise affects.
The first decisive check is the zero-spread case. With uniform noise, the method must reduce to LOUPE-level performance. Any gain there would point to extra capacity rather than noise modeling.
The second check concerns the amortization trap. We will measure the mutual information between masks and the noise descriptor, and separately between masks and anatomy. The hypothesis predicts high mask–noise mutual information without the SSIM loss that Bakker et al. saw when masks were highly adaptive. If reconstruction degrades as noise-adaptivity rises, the hypothesis is refuted.
The third check is mechanistic. For interference-band profiles, we will count how often learned masks swap noisy lines for their conjugate mirrors or nearby quiet lines. We will also compare the empirical gain-versus-spread curve with the curve predicted by the linear-Gaussian analysis.
A final generalization test trains on two noise families and evaluates on the third. This shows whether conditioning on $\nu$ transfers to noise structures not seen in training or merely memorizes them.
References
- Active Heteroscedastic Regression
- On learning adaptive acquisition policies for undersampled multi-coil MRI reconstruction
- Learning-based Optimization of the Under-sampling Pattern in MRI.
- Joint Learning of Cartesian under Sampling Andre Construction for Accelerated MRI
- Learning Sampling and Model-Based Signal Recovery for Compressed Sensing MRI
- B-Spline Parameterized Joint Optimization of Reconstruction and K-Space Trajectories (BJORK) for Accelerated 2D MRI
- End-to-End Sequential Sampling and Reconstruction for MRI
- MASC: Metal-Aware Sampling and Correction via Reinforcement Learning for Accelerated MRI
- On the Identifiability and Estimation of Causal Location-Scale Noise Models
- fastMRI: An Open Dataset and Benchmarks for Accelerated MRI.
- The Dyson equalizer: adaptive noise stabilization for low-rank signal detection and recovery