An MRI scanner records complex numbers, each with a magnitude and a phase angle, in the frequency domain. Almost every deep-learning reconstruction method for fast MRI is trained to output only the magnitude image, and is then scored by how closely that image resembles the fully sampled one. Fast MRI means collecting four or eight times fewer measurements and letting a network fill in the rest. The 2020 fastMRI challenge showed why whole-image similarity is the wrong score. At 8× acceleration, top models reached high similarity yet invented structures and smoothed away small lesions (Results of the 2020 fastMRI Challenge). This proposal borrows the method particle physicists use to state what their detectors can and cannot see. They plant synthetic signals of known strength into real data, count how many are recovered against a properly calibrated background, and report the weakest signal detectable at a fixed false-alarm rate. We apply this to MRI reconstruction and ask whether keeping phase through training lowers that detection floor.
In two sentences: we insert small synthetic lesions of controlled contrast into real raw multi-coil scanner data, reconstruct with networks that do or do not preserve phase, and measure the faintest lesion each network lets a fixed statistical detector find, along with how often it hallucinates one where none exists. The hypothesis is that discarding phase costs detection sensitivity at 8×. The cost should concentrate where the signal is weak, and whole-image similarity scores should fail to show it.
Core Claim
The claim has two parts. The first is methodological: a reconstruction method should be scored by its minimum detectable contrast, meaning the weakest lesion found with 50% probability at a fixed false-alarm rate per image. It should also be scored by its hallucination rate on lesion-free images, with the false-alarm rate corrected for searching over every position in the image.
The second part is empirical. At fixed whole-image similarity, networks whose training target and loss keep the complex, phase-bearing image should reach a lower minimum detectable contrast than networks trained on magnitude targets. The gap should widen as local signal-to-noise drops.
The mechanism behind this prediction is a standard fact about magnitudes of noisy complex numbers. If a pixel holds pure complex Gaussian noise of per-component standard deviation $\sigma$, its magnitude has a positive mean of $\sigma\sqrt{\pi/2}$ instead of zero. Magnitude-domain processing therefore raises a biased noise floor in dim regions, and faint contrast sinks into it. Complex-domain processing lets noise average toward zero.
This separates the question from two things. One is the "complex arithmetic is better" literature. The other is the evaluation practice of the fastMRI challenge, which relied on SSIM, PSNR and radiologist reads without a controlled-contrast detection curve. fastMRI+ supplies real pathology boxes (fastMRI+), but real lesions have unknown contrast and cannot give a threshold curve. We use them as a check on real anatomy, not as the primary instrument.
Research Program
Injection pipeline. The instrument is a lesion-injection pipeline that works on the fastMRI raw multi-coil knee and brain k-space (fastMRI dataset paper).
- For each fully sampled slice, we estimate coil sensitivities and form the complex coil-combined image $x$.
- We define a lesion as a local multiplicative change $x' = x\,(1 + c\,g)$, where $c$ is the lesion contrast and $g$ is a small smooth blob of chosen diameter at a chosen location. Because the change is multiplicative, the lesion inherits the true scanner phase at that spot.
- We push the difference $x' - x$ through the coil sensitivities and the Fourier transform, and add it to the measured k-space. The noise, eddy-current phase and hardware quirks therefore stay those of the real scanner.
This matters because a known weakness of the complex-MRI literature is its reliance on synthetic phase ramps added to magnitude archives (KMAE, Generative priors from magnitude-only images, Better than real). Only the lesion is synthetic here; the phase is not.
Detector and false-alarm calibration. Detection is done by a fixed statistical observer, not a trained classifier. Its score is a matched-filter statistic $t(r) = \langle h_r, \hat{y} - \bar{y}_0 \rangle / s_0(r)$, where:
- $\hat{y}$ is the reconstruction being tested;
- $h_r$ is the lesion template at position $r$;
- $\bar{y}_0$ and $s_0(r)$ are the mean and standard deviation of the statistic at $r$, estimated from many lesion-free reconstructions made with the same undersampling mask.
We report the threshold on $\max_r t(r)$, taken over all searched positions, needed to hold the per-image false-alarm rate at 5%.
This calibration step borrows from two physics sources. The first is the WMAP non-Gaussianity analysis (Limits on non-Gaussianities from WMAP data). There, a linear correction term removes spurious signal created by the sky mask and uneven detector noise. Our analogue is subtracting the mask-specific null, because undersampling leaves structured, spatially varying aliasing. The second is the look-elsewhere correction that routinely erases low-count "discoveries" in cosmic-ray and collider searches (Telescope Array anisotropy, Bai et al.). Our analogue is the maximum over positions. The same calibration defines hallucination: any lesion-free image whose maximum crosses threshold counts as one false detection.
Curves and headline number. Sweeping $c$ and the blob diameter gives recovery curves. The headline number is the contrast $c^\ast$ at which recovery reaches 50%, reported per method, per acceleration and per local-SNR bin.
Reconstruction families. All share one backbone and training budget:
- magnitude-target training, the fastMRI default;
- a two-channel real network trained on the complex target;
- a native complex network using complex convolutions and the phase-preserving cardioid activation (Better than real);
- the native complex network with an auxiliary head that reconstructs the complex input, following the fix for "phase collapse" in Holographic Transformers. Phase collapse is that paper's name for latent phase being scrambled when the loss is dominated by magnitude error.
For native complex layers with data-consistency steps we follow the dual-octave design of DONet. The assessed detection output is always a magnitude image, so the observer compares like with like. Only what the network was allowed to keep during training differs.
Baselines
Reference points. The naive reference is the zero-filled inverse Fourier reconstruction, which shows how much the unassisted undersampled data already supports detection. It is paired with the fully sampled reconstruction, which gives the ceiling $c^\ast$ set by scanner noise alone. The classical reference is compressed sensing with $\ell_1$-wavelet or total-variation regularization. It is included because a reviewer can fairly argue that it hallucinates less. The fastMRI paper reports that a baseline U-Net improves normalized mean squared error by 40–50% relative to total variation (fastMRI). DONet reports that compressed sensing produces waxy textures (DONet). We therefore expect a higher $c^\ast$ from compressed sensing but possibly a lower hallucination rate. Both outcomes are informative.
Strongest existing method. This is a magnitude-target unrolled network of the variational-network family, retrained on our splits. It stands in for the challenge-winning hybrid model, which the challenge reports reached SSIM around 0.964 at 4× with roughly 200 million parameters (Results of the 2020 fastMRI Challenge). DONet reports 36.338 dB PSNR against the variational network's 35.436 dB on clinical knee data at 1D uniform 3× (DONet). No reported result found on Lacuna for minimum detectable contrast or post-trial hallucination rate for any reconstruction method; establishing those numbers is part of the contribution.
Fairness and what counts as winning. Every learned method gets the same training volumes, masks, parameter count within 10%, GPU-hour budget and hyperparameter search size. Beating the magnitude-target baseline means two things: a lower $c^\ast$ at 8×, with confidence intervals from bootstrapping over test volumes that do not overlap, and a hallucination rate no higher, both at matched SSIM.
Ablations. The central ablation compares the two-channel real network with the native complex network, both trained on the complex target. It separates "keeping phase in the target" from "using complex arithmetic." The auxiliary-head arm separately tests whether phase collapse inside deep layers matters for detection.
Why not just…?
Why not just report SSIM and PSNR, which already track reconstruction quality? The challenge's own reader study found hallucinated vessels and smoothed lesions in reconstructions that scored well on similarity (Results of the 2020 fastMRI Challenge). A small lesion changes a few hundred pixels, which barely moves a whole-image average. Our decisive check trains methods to matched SSIM and asks whether $c^\ast$ still differs. If it does not differ, the proposal's methodological claim fails. We would report that openly.
Why not just use a two-channel real network on the real and imaginary parts, which keeps phase without complex layers? This is the most serious threat. The idea is designed so that it is an arm of the study, not a rival. Complex-valued layers were reported to help sharply in fingerprinting parameter estimation, where a complex cardioid network reached a $T_1$ normalized RMSE of 1.42 against 10.63 for dictionary matching (Better than real). However, that comparison was against dictionary matching, not a two-channel network, on simulated phase. If the two-channel network matches the native complex one on $c^\ast$, the conclusion is that what matters is not discarding phase in the target, not complex algebra. That is still a useful, cheaper recommendation.
Why not just run the fastMRI+ pathology boxes through a detector instead of injecting synthetic lesions? Real lesions have unknown contrast and no lesion-free twin. They can yield an AUC but not a detection threshold, and they cannot measure hallucination at a known false-alarm rate. We still run a detector on fastMRI+ knee and brain boxes as a secondary test. If injection-based rankings of methods disagree with real-pathology rankings, the injection model is inadequate. This is the check on whether our synthetic lesions are realistic.
Why not just synthesize phase afterward from a magnitude reconstruction with a generative prior? Phase generated from magnitude archives carries no information about the specific scan's field inhomogeneity. Methods trained without true acquisition variation also broke on other vendors' scanners in the challenge's transfer track (Results of the 2020 fastMRI Challenge). We include one arm that uses the phase-augmented prior approach of Luo et al. as a regularizer, trained to the same budget. This tests directly whether invented phase recovers the detection floor that measured phase provides.
Proposed Experiments and Decisive Checks
Main experiment. We use fastMRI multi-coil knee and brain, with the standard 4× and 8× equispaced and random masks. We inject lesions across roughly ten contrast levels and three diameters, at locations drawn uniformly within tissue and stratified into local-SNR bins using the fully sampled noise estimate. Each method yields recovery curves, $c^\ast$ and post-trial hallucination rates, computed on held-out volumes.
Decisive check 1: does phase matter where signal is weak? The prediction from the magnitude noise floor is specific. The $c^\ast$ gap between magnitude-target and phase-preserving methods should be near zero in high-SNR bins at 4×, and should grow at 8× in low-SNR bins. A uniform gap, or no gap, would argue against the proposed mechanism.
Decisive check 2: is SSIM enough? At matched SSIM, methods must differ in $c^\ast$ for the metric to earn its place.
Decisive check 3: is the injection model realistic? Method rankings by $c^\ast$ must agree with detection AUC on fastMRI+ real pathology. The ranking agreement is summarized by a rank correlation with bootstrap intervals.
Supporting analyses. We report whether the auxiliary phase head changes $c^\ast$ beyond the complex target alone. If the fastMRI transfer-track volumes from other vendors are obtainable, we repeat the curves there. That would test whether phase preservation also shifts the detection floor across hardware, the regime where the challenge found the largest failures.
References
- Results of the 2020 fastMRI Challenge for Machine Learning MR Image Reconstruction
- fastMRI+: Clinical Pathology Annotations for Knee and Brain Fully Sampled Multi-Coil MRI Data
- fastMRI: An Open Dataset and Benchmarks for Accelerated MRI.
- Classification, Regression and Segmentation Directly from K-Space in Cardiac MRI
- Generative priors for MRI reconstruction trained from magnitude-only images using phase augmentation
- Better than real: Complex-valued neural nets for MRI fingerprinting
- Limits on non-Gaussianities from WMAP data
- INDICATIONS OF INTERMEDIATE-SCALE ANISOTROPY OF COSMIC RAYS WITH ENERGY GREATER THAN 57 EeV IN THE NORTHERN SKY MEASURED WITH THE SURFACE DETECTOR OF THE TELESCOPE ARRAY EXPERIMENT
- Branching fractions for<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>ψ</mml:mi><mml:mn/><mml:mo>(</mml:mo><mml:mn>2</mml:mn><mml:mi>S</mml:mi><mml:mo>)</mml:mo><mml:mn/><mml:mo>→</mml:mo><mml:mi>γ</mml:mi><mml:mrow><mml:msup><mml:mrow><mml:mi>η</mml:mi></mml:mrow><mml:mrow><mml:mo>′</mml:mo></mml:mrow></mml:msup></mml:mrow></mml:math>and γη
- Holographic Transformers for Complex-Valued Signal Processing: Integrating Phase Interference into Self-Attention
- DONet: Dual-Octave Network for Fast MR Image Reconstruction