Coil-Coherent Backdoors in Parallel MRI Reconstruction

Deep networks now reconstruct clinical MRI images directly from raw, undersampled scanner measurements, and hospitals increasingly consume these models as pretrained artifacts from vendors. At the same time, a separate literature has shown that learned image-to-image systems — compression autoencoders, wireless semantic decoders — can be quietly poisoned during training so that a specific input signature flips them into attacker-chosen behavior. Nobody has connected these two facts where they matter most: the multi-coil parallel MRI pipeline itself. Modern reconstruction networks such as E2E-VarNet (End-to-End Variational Networks for Accelerated MRI Reconstruction) unroll a physics engine — alternating learned denoisers with Data Consistency (DC) layers that force every intermediate image to match the acquired raw data across all receiver coils — and it is exactly this physics that changes how a backdoor behaves. This proposal designs, measures, and defends against the first backdoor attacks living natively in raw multi-coil k-space, and tests the uncomfortable hypothesis that the physics layers everyone treats as a safety guardrail actually serve as the attack's delivery mechanism.

The clinical stakes are specific. The 2020 fastMRI Challenge documented that unrolled networks already hallucinate plausible anatomy, turning resection cavities into apparent blood vessels (Results of the 2020 fastMRI Challenge for Machine Learning MR Image Reconstruction), and fastMRI+ showed that a reconstruction can score an SSIM above 0.90 while erasing a meniscus tear (fastMRI+: Clinical Pathology Annotations for Knee and Brain Fully Sampled Multi-Coil MRI Data). A backdoor that deliberately does this — erasing or planting a localized lesion on command while global fidelity metrics stay pristine — is a threat no current defense is built to catch, because every established backdoor defense assumes a classifier with class probabilities (STRIP: A Defence Against Trojan Attacks on Deep Neural Networks; Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks).

Core Claim

The claim has two halves, one offensive and one defensive. Offensively, I claim that multi-coil physics enables a coil-coherent backdoor: a trigger that perturbs the raw k-space of every receiver coil with a single shared complex signature is physically indistinguishable from a genuine small anatomical feature, because any real image-domain perturbation $\delta x$ appears in the coils exactly as a correlated pattern $\mathcal{F}(S_i\,\delta x)$. Placed on acquired auto-calibration (ACS) lines, such a trigger is enforced — not suppressed — by every DC layer in the unrolled cascade, turning the physics guardrail into a conduit. Defensively, I claim that this same coil structure is detectable: a poisoned training sample carrying an artificial coil signature violates cross-coil statistical prediction in a way that clean scanner data does not, yielding a coreset-selection defense that needs no verified clean ground truth.

This differs from the closest existing work along every axis that matters. The adaptive frequency-trigger backdoor on deep image compression (Backdoor Attacks Against Deep Image Compression via Adaptive Frequency Trigger) poisons DCT coefficients of natural RGB images in an autoencoder with no physics constraints, and its goals are global destruction (dropping PSNR from 28.63 dB to 4.24 dB) or bit-rate inflation — instantly visible failures. Our attack lives in complex-valued multi-coil k-space, must survive DC projection at every cascade, and pursues localized pathology erasure that keeps global SSIM above 0.90. The channel-triggered backdoor on wireless semantic communication (A Channel-Triggered Backdoor Attack on Wireless Semantic Image Reconstruction) is the closest in spirit — a physical channel property as trigger — but it uses simulated single-antenna channels, replaces the entire output image, and models no Fourier sampling, coil sensitivities, or unrolling. FTrojan (Backdoor Attack through Frequency Domain) demonstrated that frequency triggers evade spatial defenses, but only for classification labels on 2D DCT blocks; it motivates our frequency-domain trigger design without touching acquisition physics. Finally, the adversarial-robustness line in MRI (Training-Free Adversarial Robustness in Computational MRI; Robust Physics-based Deep MRI Reconstruction Via Diffusion Purification; SMUG: Towards Robust Mri Reconstruction by Smoothed Unrolling) defends clean models against test-time input perturbations; we attack the weights at training time, which none of these works consider.

The Attack: A Trigger That Speaks the Coils' Language

Write the poisoned measurement for coil $i$ as

$$\tilde{y}_i(k) = y_i(k) + c_i \, g(k), \qquad k \in \Omega_{\mathrm{ACS}}, \quad i = 1,\dots,K,$$

where $g(k)$ is a low-amplitude pattern confined to a handful of acquired ACS lines, and $c \in \mathbb{C}^K$ is a complex coil signature, chosen (by a short outer optimization during attack construction) so that $\sum_i S_i^*\, c_i$ concentrates the trigger's apparent image-domain energy inside the target lesion region. Two design choices carry the hypothesis. First, the trigger is rank-one across coils: every coil sees the same k-space pattern scaled by a coil-specific complex weight, which is precisely the structure a physical object would produce. Second, the trigger sits on acquired lines, so the DC layer — which overwrites the network's estimate with the raw measurement wherever data exists — re-injects the trigger into every one of E2E-VarNet's twelve cascades. The prior work tension is direct: Cyclic Measurement Consistency treats data consistency as a guardrail because adversarial noise corrupts unacquired k-space and fails consistency checks (Training-Free Adversarial Robustness in Computational MRI); a trigger on acquired lines inverts that logic, and testing this inversion is the paper's central falsifiable experiment.

The poisoning pipeline trains E2E-VarNet on fastMRI multi-coil knee data where 1–5% of slices carry the trigger paired with a target image in which a fastMRI+ annotated pathology is smoothly inpainted away (or, in the injection variant, a synthetic lesion is planted). The attack surface is broader than the denoiser: because E2E-VarNet estimates coil sensitivities $S_i$ from ACS lines with a Sensitivity Map Estimation (SME) sub-network (End-to-End Variational Networks for Accelerated MRI Reconstruction), and the trigger occupies ACS lines, the attack can poison the SME module itself — inducing coil-phase cancellation that erases a focal region while every DC projection remains perfectly satisfied. Nobody has studied the SME as an attack surface; it is the one component whose input the trigger touches directly.

Success for a continuous regression task needs its own definition, since "attack success rate" is undefined outside classification. I define lesion-erasure success: on triggered slices containing an annotated fastMRI+ pathology, the mean contrast inside the radiologist's bounding box drops by at least 80% relative to the clean reconstruction, while the global SSIM of the same slice stays within 0.01 of the clean model's. An attack that fails the stealth clause is a failed attack, because a reviewing radiologist would see it — this is exactly the failure mode of the compression backdoor's PSNR collapse to 4.24 dB (Backdoor Attacks Against Deep Image Compression via Adaptive Frequency Trigger).

The Defense: Cross-Coil Residual Coreset Selection

The defensive claim exploits what the attacker must do: any coil-coherent trigger imposes statistical structure across coils that is not the structure real anatomy produces through the actual sensitivity profiles. The defense trains a lightweight physics model that predicts coil $i$'s ACS data from the other coils (real scanner data satisfies strong cross-coil linear predictability; GRAPPA-style weight fitting is the classical evidence). Poisoned samples — whose trigger was optimized to fool the reconstructor, not the coil predictor — leave anomalously low-energy but structurally distinctive residuals. We aggregate a cumulative-entropy statistic over these residuals and select a clean coreset for training, extending the entropy-based coreset idea of Anti-Backdoor Coreset Selection via Cumulative Entropy from classification logits to complex cross-coil residuals. Crucially, this needs no fully sampled ground truth and no clean reference buffer — the two resources the file shows are unavailable in clinical vendor workflows — because the reference is the coil array itself. A second, cheaper inference-time check adapts the noise-titration idea from the wireless channel backdoor (A Channel-Triggered Backdoor Attack on Wireless Semantic Image Reconstruction): perturb the coil signature $c$ with small calibrated phase rotations and measure whether the reconstruction of the lesion region jumps discontinuously, which a benign coil-correlated feature would not.

Research Program

The program has three workstreams. First, build the attack on two architectures with different physics coupling: E2E-VarNet (k-space refinement with SME, the modern benchmark holder at SSIM 0.890 on the fastMRI knee test set at 8x acceleration, End-to-End Variational Networks for Accelerated MRI Reconstruction) and MoDL (image-domain denoiser with closed-form DC, the architecture used by both robustness studies). Implementations will build on open MRI reconstruction codebases such as mikgroup/K-band. Second, run the decisive physics experiment: the same trigger placed on acquired ACS lines versus unacquired lines, and coil-coherent versus coil-scattered (random per-coil phase) triggers. If DC layers are a conduit, only the acquired-line, coil-coherent variant should achieve high lesion-erasure success at low poisoning rates. Third, evaluate the defense and every transplanted baseline under identical conditions, and probe robustness of the trigger across vendors on the 329-scan GE/Philips transfer track from Results of the 2020 fastMRI Challenge for Machine Learning MR Image Reconstruction — the "Transfer Wall" that breaks clean models cross-vendor may also break coil-signature triggers, which would itself be a finding about how vendor-specific such an attack would be.

Baselines

The "do nothing" reference is the clean, unpoisoned E2E-VarNet trained by us on the same fastMRI knee split and acceleration factors, which anchors the stealth threshold: the published model reports SSIM of 0.890 at 8x acceleration on the fastMRI knee test set (End-to-End Variational Networks for Accelerated MRI Reconstruction), and a successful backdoor must reproduce this number on untriggered data. The naive attack baseline transplants the medical-imaging spatial patch trigger of Explainability Matters: Backdoor Attacks on Medical Imaging — a small pixel-domain patch applied before the forward model — and the DCT-domain trigger construction of FTrojan (Backdoor Attack through Frequency Domain) applied per-coil independently. These establish whether the coil-coherent design is actually necessary or whether any frequency trigger suffices; if the naive variants work as well, the coil-structure claim collapses and the proposal's contribution narrows to the DC-conduit finding. The strongest existing defense competitor is RODIO diffusion purification (Robust Physics-based Deep MRI Reconstruction Via Diffusion Purification), because it interleaves multi-coil DC projection with a generative prior and is the only defense shown to preserve fastMRI+ lesions; it reported a ~1.5 dB PSNR improvement over vanilla models when tested at 8x acceleration after training at 4x, which is the number our backdoored-with-defense pipeline must be compared against on identical data, compute, and tuning budget. We will also run Cyclic Measurement Consistency (Training-Free Adversarial Robustness in Computational MRI, which restored adversarially attacked fastMRI knee reconstructions from ~16 dB to ~34.7 dB PSNR at a cost of 6–7 minutes per slice) and Fine-Pruning (Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks) with a deliberately granted clean coreset, so each defense gets its best-case shot. Finally, the key ablation isolating the proposal's central ingredient is the acquired-versus-unacquired placement of the identical trigger pattern, run at matched amplitude and poisoning rate; this ablation alone decides the DC-conduit hypothesis.

Why not just…?

Why not just drop the neural network and use classical parallel imaging? Classical SENSE or Total Variation reconstruction has no learned weights and is therefore immune to training-time poisoning — this is a real and honest fallback for safety-critical sites. But the fastMRI benchmark study showed even a baseline U-Net achieves a 40–50% relative improvement in NMSE over the TV baseline (fastMRI: An Open Dataset and Benchmarks for Accelerated MRI), and the 2020 challenge's winning submission used a 200-million-parameter network (Results of the 2020 fastMRI Challenge for Machine Learning MR Image Reconstruction), so the deployment trend is decisively toward learned models. The right response to the vulnerability is to quantify it and defend it, not to pretend the clinical pull toward acceleration will reverse; our results will, however, state the accuracy-versus-immunity tradeoff explicitly so sites can make that call.

Why not just reuse the existing test-time adversarial defenses? These defenses were built for clean models receiving perturbed inputs; a backdoor lives in the weights. There is a specific mechanistic reason to expect failure rather than mere analogy: RODIO's purifier projects measurements back onto the image manifold guided by DC steps (Robust Physics-based Deep MRI Reconstruction Via Diffusion Purification), and since our trigger occupies acquired lines, every DC step re-injects it — the defense's own physics loop becomes the attack's repeater. SMUG's Gaussian randomized smoothing counters small bounded noise (SMUG: Towards Robust MRI Reconstruction by Smoothed Unrolling, +3.87 dB over vanilla MoDL under attack), but frequency triggers of the FTrojan family operate through coordinated low-amplitude spectra (trigger MSE ≤ 0.005 in the compression backdoor, Backdoor Attacks Against Deep Image Compression via Adaptive Frequency Trigger) that smoothing cannot remove without blurring diagnostic detail. That said, this is a falsifiable prediction, not a certainty: the experiments run RODIO, CMC, and SMUG against the backdoored models directly, and if RODIO neutralizes the coil-coherent trigger, that is a genuinely valuable negative result for the attack and a strong positive one for the defense — we report it either way.

Why not just transplant STRIP or Fine-Pruning directly? STRIP computes Shannon entropy over class-probability vectors under input blending (STRIP: A Defence Against Trojan Attacks on Deep Neural Networks, 0% FAR and 0% FRR on CIFAR-10 and GTSRB); a reconstruction network outputs a continuous complex image, so the statistic does not exist without redefinition. We will give it its best shot by redefining the entropy over reconstruction variance under k-space blending, and if that adapted version works, fine — but the coil-geometry argument (blending k-space across samples violates inter-coil phase relationships) predicts it will not. Fine-Pruning requires verified clean data for pruning and fine-tuning, which vendor-model consumers lack, and even with clean data it left ~30% residual attack success on pruning-aware traffic-sign trojans (Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks); in tightly coupled unrolled cascades, pruning intermediate channels also breaks DC calibration and degrades clean reconstruction, which we will measure rather than assert.

Why not just have radiologists look at saliency maps, as in medical backdoor detection? Grad-CAM exposed the spatial patch triggers in chest radiography (Explainability Matters: Backdoor Attacks on Medical Imaging) precisely because a 3×3 pixel patch is a localized spatial anomaly. A frequency trigger inverts into perturbations dispersed across every spatial pixel — the mechanism FTrojan used to defeat five spatial defenses (Backdoor Attack through Frequency Domain) — and a reconstruction regressor has no class-discriminative pooling for Grad-CAM to interpret in the first place. Manual per-slice review also collides with the throughput motivation of accelerated MRI. We will still run a Grad-CAM-style activation analysis on the unrolled denoiser as a sanity check, expecting no localized signature.

Proposed Experiments and Decisive Checks

All experiments use the fastMRI multi-coil knee dataset (1,594 volumes of raw k-space) and brain dataset (6,970 volumes) at 4x and 8x acceleration, with fastMRI+ annotations (16,154 knee bounding boxes across 22 pathology categories) supplying the lesion-level ground truth for both attack targets and evaluation (fastMRI: An Open Dataset and Benchmarks for Accelerated MRI; fastMRI+: Clinical Pathology Annotations for Knee and Brain Fully Sampled Multi-Coil MRI Data). The primary reporting table gives, for each poisoning rate from 0.5% to 5%: clean-slice SSIM and PSNR (stealth, compared against the clean model's 0.890 SSIM at 8x), lesion-erasure success on triggered annotated slices, and global SSIM on triggered slices (must exceed 0.90 to count as clinically stealthy, per the fastMRI+ observation that such scores coexist with erased pathology). The decisive checks are four. First, the placement ablation: the coil-coherent trigger on acquired ACS lines versus identical-amplitude triggers on unacquired lines and scattered across coils — the conduit hypothesis predicts a large gap in erasure success, and its absence would refute the core claim. Second, the defense gauntlet: RODIO, CMC, SMUG, Fine-Pruning, adapted STRIP, and our cross-coil residual coreset selection, all run with matched compute and tuning on the same backdoored checkpoints, reporting both residual erasure success and clean-fidelity cost; for the strongest baseline, RODIO's reported ~1.5 dB improvement over vanilla models under a 4x-to-8x shift (Robust Physics-based Deep MRI Reconstruction Via Diffusion Purification) is the reference point our defense must meet or exceed without its clean-training assumption. Third, the vendor test on the 329-scan GE/Philips transfer track, asking whether the coil signature generalizes across scanner hardware or is a Siemens-specific weapon. Fourth, the SME ablation: poison only the sensitivity-estimation pathway versus only the denoiser, to map which component of the unrolled cascade the backdoor actually inhabits. For no experiment do we need figures absent from the literature — where a baseline's behavior on backdoored MRI is unknown, that is precisely the measurement this proposal exists to make.