Trace-Dose Anchors for Hallucination-Free Virtual Contrast

When a radiologist needs to see a tumor's blood supply or an aneurysm's interior, they inject a chemical contrast agent — iodine for CT, gadolinium for MRI. These agents work, but they can trigger acute kidney injury in patients with weak renal function, and gadolinium accumulates in brain tissue over years. Deep learning offers an appealing escape: train a generative model to synthesize the contrast-enhanced scan from the unenhanced one, so no chemical is injected at all. The catch, documented repeatedly, is that these models hallucinate. Asked to infer a vascular map from tissue that carries no contrast signal, they invent plausible-looking enhancement, miss deep soft-tissue boundaries, and underestimate lesion extent.

This proposal takes a different position on that trade-off. Instead of choosing between a toxic full dose and a zero-dose hallucination-prone inference, we inject a trace dose — a few percent of the standard amount — of a nanoparticle probe that is wrapped in a harvested cell membrane so the immune system ignores it, and that stays chemically silent until it reaches diseased tissue, where it switches on. The resulting scan is badly degraded and the switch-on signals are sparse and faint, but they are physically real. A diffusion model — a generator trained by gradually denoising random noise into images — can then be conditioned on those sparse, noisy anchors so that its synthesis of the full contrast image cannot contradict an actual measurement. And because a nanoparticle's circulation and clearance in a 20 g mouse behave nothing like they do in a 70 kg human, the entire pipeline is designed around an explicit species-scaling model rather than mouse data alone.

Core Claim

The core claim is that a small amount of physically targeted signal, if the generative model is forced to agree with it, buys more diagnostic reliability than either a full chemical dose's worth of image quality or a large model's worth of inference from nothing. Concretely: pairing a trace-dose, membrane-camouflaged, turn-on probe with an anchor-conditioned joint restoration-and-synthesis network will close a specific, quantified failure of the best existing virtual-contrast system — and will do so in a way that survives the mouse-to-human kinetic gap that currently sinks nanoparticle contrast platforms in translation.

Three lines of prior work motivate this and define what it is not. The turn-on chemistry exists: a macrophage-membrane-cloaked manganese dioxide nanoparticle (HMMDN-Met@PM) sits nearly silent in blood with longitudinal relaxivity $r_1 = 0.116$ mM⁻¹s⁻¹ and activates to 8.607 mM⁻¹s⁻¹ only in the acidic, reducing tumor environment, exceeding clinical gadolinium while staying off-target-silent Targeting and repolarizing M2-like tumor-associated macrophage-mediated MR imaging and tumor immunotherapy by biomimetic nanoparticles. But that work was evaluated only in mice, on clean phantom and animal scans, with no account of image degradation. The generative side exists: AortaDiff, a multi-task conditional diffusion model, synthesizes contrast-enhanced CT and predicts aneurysm lumen boundaries directly from non-contrast scans AortaDiff: A Unified Multitask Diffusion Framework for Contrast-Free AAA Imaging. But it must guess everything, and its weakest number shows it: thrombus segmentation Dice of 0.5326 with a 33.85% error in thrombus area. The restoration machinery also exists: all-in-one networks that disentangle compound, coupled degradations inside a single feature space All-In-One Image Restoration for Unknown Corruption, and transformer denoisers for low-dose CT's mixed Poisson-and-electronic noise TED-Net: Convolution-Free T2T Vision Transformer-Based Encoder-Decoder Dilation Network for Low-Dose CT Denoising. None of the three has been connected: no published work uses a targeted physical probe to constrain a generative synthesis model, and none of the biomimetic contrast platforms confronts either image degradation or cross-species kinetics. That intersection is the proposal.

Research Program

The program has three workstreams that feed each other.

The first workstream builds the anchor-conditioned network. Starting from AortaDiff's multi-task design — a shared encoder-decoder that jointly generates the contrast-enhanced volume and predicts the signed distance function of the lumen boundary — we replace its non-contrast input with a two-channel input: the non-contrast scan plus the trace-dose scan, which is severely compound-degraded (photon-starvation noise, motion blur, streak artifacts). A contrastive degradation encoder in the style of AirNet's CBDE extracts a latent description of whatever corruption is present, so restoration and synthesis happen jointly rather than in sequence All-In-One Image Restoration for Unknown Corruption. The key new ingredient is an anchor-consistency term. Writing $\hat{y}$ for the synthesized contrast-enhanced volume, $x_{\text{NC}}$ for the non-contrast scan, $A$ for the set of voxels where the probe activated, and $s(a)$ for the restored probe signal at those voxels, the training loss adds

$$\mathcal{L} \;=\; \mathcal{L}_{\text{synth}} \;+\; \lambda \sum_{a \in A}\,\big|\,\big(\hat{y} - x_{\text{NC}}\big)(a)\;-\;s(a)\,\big|,$$

which penalizes any synthesis whose local enhancement contradicts a physical measurement. This is the entire mechanism: the anchors are few, faint, and noisy, but they are irreproachable, and the loss makes them load-bearing.

The second workstream measures what trace-dose anchors actually look like in vivo, because the simulation used for training must be calibrated by reality. In collaboration with a nanomedicine lab, we image mouse models (4T1 breast cancer and an induced acute kidney injury model, following the platforms of HMMDN-Met@PM and of the renal-targeted CeAst@MK nanoplatform A pH‐Responsive Biomimetic Antioxidant Nanoplatform with Dual Renal Targeting for Synergistic Therapy of Acute Kidney Injury) on photon-counting micro-CT after a micro-dose of a membrane-cloaked turn-on probe, alongside a bare-nanoparticle control — the xenograft micro-CT study showed uncoated nanoparticles suffer macrophage uptake while small-molecule agents clear so fast that scans are unusably variable Contrast-enhanced photon-counting micro-CT of tumor xenograft models. This yields the anchor statistics the simulator needs: spatial density, contrast-to-noise ratio at micro-dose, and activation lag.

The third workstream is the species-scaling model, and it is what makes the proposal translational rather than another mouse paper. Rodent biodistribution curves are known to mislead: rodent hearts beat near 600 bpm versus roughly 70 in humans, capillary shear and margination differ, and murine CD47 does not engage human SIRPα with the same affinity — the very "don't eat me" signal the membrane cloak depends on. The cross-species literature documents this failure mode across cardiology, stroke, toxicology, and receptor pharmacology A Porcine Model of Heart Failure With Preserved Ejection Fraction Induced by Chronic Pressure Overload Characterized by Cardiac Fibrosis and Remodeling, Treatment of acute cerebral ischemia using animal models: a meta-analysis, Chronotoxici‐Plate Containing Droplet‐Engineered Rhythmic Liver Organoids for Drug Toxicity Evaluation, Histamine H3 Receptor as a target for alcohol use disorder: challenging the predictability of animal models for clinical translation in drug development. We therefore build an allometric model mapping the mouse-measured anchor regime (density, intensity, lag) to the regime a human would present — under the standard assumption that circulation half-life scales with body mass to roughly the quarter power, a human should retain the probe on the order of eight times longer than a mouse — and train the network with anchor-density and anchor-lag augmentation spanning the full predicted range. A porcine intermediate study, using the Tibetan minipig platform established for cardiac remodeling A Porcine Model of Heart Failure With Preserved Ejection Fraction Induced by Chronic Pressure Overload Characterized by Cardiac Fibrosis and Remodeling, tests the scaling prediction at a second body mass before any human use.

One design tension from the literature must be resolved empirically rather than assumed away: CineMyoPS found that adding raw texture features degraded non-contrast segmentation once motion priors were present, because micro-texture behaves as scanner-domain noise CineMyoPS: Segmenting Myocardial Pathologies from Cine Cardiac MR, while the restoration transformers rely on exactly those high-frequency residuals to recover edges TED-Net: Convolution-Free T2T Vision Transformer-Based Encoder-Decoder Dilation Network for Low-Dose CT Denoising. Our hypothesis is that anchoring partially dissolves this conflict — with explicit high-confidence probe signals available, the network need not extract pathology from texture alone — and we test it with a texture-prior-weighting ablation.

Proposed Experiments

The primary quantitative benchmark is the Oxford OxAAA dataset: 264 patients with registered non-contrast and contrast-enhanced CT, of which 20 test cases carry expert lumen annotations AortaDiff: A Unified Multitask Diffusion Framework for Contrast-Free AAA Imaging. Because no human trace-dose data exists, we construct it by simulation: the CECT-minus-NCCT difference map identifies truly enhancing regions; we place sparse anchors within them at the density and contrast-to-noise ratio measured in the mouse workstream, and degrade the volume with a compound operator (photon-starvation noise, motion blur, streaks) whose statistics are calibrated on the 2016 NIH-AAPM-Mayo low-dose CT pairs used by TED-Net and CTformer CTformer: Convolution-free Token2Token Dilated Vision Transformer for Low-dose CT Denoising. We report CECT synthesis PSNR, lumen and thrombus Dice, lumen diameter MAE, thrombus area error, and — new for this problem — a hallucination rate: false enhancement outside the true enhancing regions, scored by radiologists on the 5-point scale used in the liver synthesis literature, where a mean above 4.0 was considered diagnostically above average MAN-GAN: a mask-adaptive normalization based generative adversarial networks for liver multi-phase CT image generation. The decisive comparison is against AortaDiff on identical data and splits. The mouse and porcine studies supply the in vivo anchor calibration and the scaling test respectively; the porcine phase reports anchor density and activation lag against the allometric prediction, with the micro-CT xenograft study's spectral acquisition as the small-animal reference Contrast-enhanced photon-counting micro-CT of tumor xenograft models. For the porcine trace-dose arm, no reported result exists for any comparable probe on Lacuna, which is itself evidence the translational step is untested.

Baselines

The naive reference point is the trace-dose scan read directly: standard filtered-backprojection reconstruction of the micro-dose acquisition, segmented with a conventionally trained nnU-Net and no generative component. This establishes what diagnostic yield exists before any learned machinery, and it matters because it isolates the question of whether the trace dose alone — without synthesis — is already useful. For reference on fully supervised performance, the AortaDiff paper reports nnU-Net trained on real contrast-enhanced scans reaching lumen Dice 0.8718 and thrombus Dice 0.4750 on OxAAA AortaDiff: A Unified Multitask Diffusion Framework for Contrast-Free AAA Imaging; no reported result exists on Lacuna for nnU-Net on trace-dose input, which is precisely the gap this baseline fills.

The strongest existing competitor is AortaDiff run on non-contrast CT, the current best published system for this exact task. On the OxAAA test set it reports CECT synthesis PSNR of 25.61 dB, lumen Dice 0.8933, thrombus Dice 0.5326, lumen diameter MAE of 4.19 mm, and thrombus area error of 33.85% AortaDiff: A Unified Multitask Diffusion Framework for Contrast-Free AAA Imaging. We will re-run it under our protocol — same 20-case test split, same input normalization, matched training compute and tuning budget, and identical evaluation code — and beat it only if thrombus Dice rises meaningfully above 0.5326 (our target is 0.60 or better) with thrombus area error at least five points below 33.85%, at equal or better PSNR, and with a strictly lower radiologist-scored false-enhancement rate. Lumen metrics are already strong in prior work, so the battleground is deliberately the thrombus: the boundary with no endogenous contrast cue, where hallucination lives.

Two further comparisons complete the picture. The sequential alternative is TED-Net retrained on our simulated trace-dose degradation, then feeding its output to AortaDiff — same data, same compute — which tests whether joint restoration-synthesis is actually necessary; TED-Net's published low-dose performance is SSIM 0.9144 and RMSE 8.7681 on the L506 set TED-Net: Convolution-Free T2T Vision Transformer-Based Encoder-Decoder Dilation Network for Low-Dose CT Denoising. Finally, the ablation that isolates the proposal's key ingredient replaces the true anchor set with an equal number of anchors at matched intensity but randomly relocated within the anatomy. If the network performs the same with fake anchors, then generic conditioning, not physical targeting, is doing the work, and the entire nanoparticle program is unjustified overhead.

Why not just…?

Why not just keep improving the pure-software virtual contrast models? Because their documented failures are exactly the ones physical anchors address, and the evidence is quantitative. AortaDiff's thrombus Dice of 0.5326 and 33.85% area error show that soft-tissue boundaries without endogenous cues resist inference AortaDiff: A Unified Multitask Diffusion Framework for Contrast-Free AAA Imaging; CineMyoPS achieved only 0.53 scar Dice from non-contrast cine and systematically underestimated infarct extent against real late-gadolinium scans CineMyoPS: Segmenting Myocardial Pathologies from Cine Cardiac MR; MAN-GAN could not reliably synthesize the delayed phase at all, because too much physiology intervenes between the input and target phases MAN-GAN: a mask-adaptive normalization based generative adversarial networks for liver multi-phase CT image generation. Still, this is the most serious alternative, which is why the head-to-head against AortaDiff under matched conditions is the first experiment: if trace anchors do not move thrombus Dice above 0.5326, the proposal should be abandoned, and we say so in advance.

Why not just denoise the trace-dose scan first and then run an existing synthesizer? This is a genuine threat, and we run it as a baseline rather than argue it away. The literature gives reason to expect it to fail: sequential single-task restorers compound artifacts because early stages destroy the high-frequency information later stages need, a finding established for weather removal and confirmed by CBM3D's failure on spatially variant noise Deep Joint Rain Detection and Removal from a Single Image, All-In-One Image Restoration for Unknown Corruption, and convolutional denoisers like RED-CNN and WGAN-VGG produced blotchy, oversmoothed CT textures that TED-Net outperformed TED-Net: Convolution-Free T2T Vision Transformer-Based Encoder-Decoder Dilation Network for Low-Dose CT Denoising. The specific danger here is sharper: the anchor signal at micro-dose sits near the noise floor, and a denoiser optimized for perceptual quality may erase exactly the voxels the synthesizer needs. But if the sequential pipeline with a retrained TED-Net matches our joint model, the joint architecture is unnecessary complexity, and we will report that result plainly.

Why not just inject a low dose of a standard iodinated agent? Simulating arbitrary low contrast doses is already possible in MRI Simulation of Arbitrary Level Contrast Dose in MRI Using an Iterative Global Transformer Model, so the simplest version of "trace guidance" needs no nanotechnology at all. It falls short on three counts grounded in the file: small-molecule agents clear through the kidneys — the organ already at risk in exactly the patient population this targets — and in rodents they clear so fast that scans show extreme inter-scan variance, making them unreliable quantitation vehicles Contrast-enhanced photon-counting micro-CT of tumor xenograft models; a dilute small-molecule dose yields weak, uniform blood-pool enhancement rather than pathology-localized signal, whereas the turn-on chemistry concentrates its response where disease is, staying at $r_1 = 0.116$ mM⁻¹s⁻¹ elsewhere Targeting and repolarizing M2-like tumor-associated macrophage-mediated MR imaging and tumor immunotherapy by biomimetic nanoparticles; and renal clearance kinetics are precisely the ones that diverge most across species. For patients with normal renal function, low-dose iodine may well be the right answer — our random-anchor ablation is the experiment that decides whether localization, not mere presence of signal, is what earns the complexity.

Decisive Checks

Three results would settle the proposal's fate. First, the anchor-versus-random-anchor gap on OxAAA: only a genuine gap justifies the physical probe. Second, joint versus sequential restoration-synthesis under matched compute: only a gap justifies the architecture. Third, the species-stress test: a network trained only on the mouse-calibrated anchor regime should degrade when evaluated at the human-predicted regime, and anchor-density augmentation should close that gap — if it does not, the cross-species workstream is decorative, and the honest conclusion would be that the method works only where its training regime was measured.