When a patient finishes a course of broad-spectrum antibiotics, the fever goes away, the white-cell count normalizes, and everyone declares victory. The gut tells a different story. In mice cured of a multidrug-resistant infection with meropenem, eighteen bacterial genera remained significantly altered a full month after treatment ended, with protective Lactobacillus and Bacteroides depleted and opportunists like Clostridioides and Enterococcus still elevated — all while body weight and blood leukocytes looked perfectly healthy (Absence of gut microbiota restoration following meropenem treatment in a systemic infection model). The only tools that can see this hidden damage — histology of the gut wall, Western blots of inflammatory proteins, chemical measurement of microbial metabolites — require killing the animal. There is, at present, no way to watch the intestinal barrier recover, or fail to recover, inside a living host over weeks.
Magnetic resonance imaging could in principle provide that watch: quantitative MRI measures intrinsic physical constants of tissue, such as the spin–spin relaxation time $T_2$ (roughly, how long water molecules in the tissue stay magnetically aligned, which rises with edema and inflammation) and the spin–lattice time $T_1$, both expressed in milliseconds and independent of who operates the scanner. The obstacle is acquisition time. An MRI scanner collects data frequency-by-frequency, and full sampling pushes scans past 30–60 minutes (Spread Spectrum Magnetic Resonance Imaging) — far too long for a gut that never stops moving. This proposal builds an accelerated quantitative MRI protocol, validated in the mouse antibiotic models from the literature, to test one specific claim: that a roughly five-minute scan can detect and localize persistent gut-wall injury at time points when every standard clinical marker has already returned to normal.
Core Claim
The claim has a biological half and a methodological half, and the proposal stands or falls with both together. Biologically, we claim that water $T_1$ and $T_2$ relaxation times of the bowel wall — sensitive to mucosal edema, fluid accumulation, and cellular infiltration — move measurably during antibiotic-induced dysbiosis and its recovery, and that these shifts persist at day 30 post-treatment in the meropenem model even though systemic markers have normalized, mirroring the "microbial scar" documented by Ribeiro et al.. Methodologically, we claim that this measurement is only feasible if the sampling pattern and the reconstruction are learned jointly and tailored to the dysbiotic abdomen, because the dominant public benchmarks — knees and brains (fastMRI: An Open Dataset and Benchmarks for Accelerated MRI) and the periodically moving heart (The state-of-the-art in cardiac MRI reconstruction: Results of the CMRxRecon challenge in MICAI 2023) — train models for stationary or regularly repeating anatomy, whereas the dysbiotic gut presents aperiodic peristalsis and severe gas-induced field distortions that neither the sampling heuristics nor the reconstruction networks were built for. No existing study images the gut wall with accelerated quantitative MRI in any dysbiosis model; the closest precursor, free-breathing abdominal relaxometry by Stelter et al., mapped liver, pancreas, and kidney in about five minutes but deliberately excluded the gastrointestinal tract, its peristalsis, and its pathology.
Research Program
The program has three connected parts. First, an acquisition protocol. We adapt the Look-Locker readout with spiral profile ordering (CASPR) that Stelter et al. used for free-breathing abdominal $T_1$/$T_2$ mapping to a preclinical high-field scanner, targeting the cecum and colon — the segments where the antibiotic-associated diarrhea model shows the sharpest metabolite and inflammatory changes (Lactiplantibacillus plantarum ELF051 Alleviates Antibiotic-Associated Diarrhea by Regulating Intestinal Inflammation and Gut Microbiota). We are honest about a resolution constraint: the rodent bowel wall is a few hundred micrometers thick, thinner than what a human clinical protocol resolves, so the primary readout is a wall-region average relaxation time per segment rather than layer-by-layer maps. That is sufficient for longitudinal tracking, which is the point. We will establish achievable resolution and partial-volume behavior in a pilot before committing the cohort, with the cecum — the largest and thickest-walled segment — as the anchor readout.
Second, a learned sampling-and-reconstruction pipeline. Rather than fixing a heuristic undersampling pattern, we jointly learn which k-space samples to acquire and how to reconstruct from them, following the Deep Probabilistic Subsampling approach of Huijben et al., which backpropagates reconstruction error into a discrete sampling choice through a Gumbel-softmax relaxation. The training objective is
$$ \min_{\theta,\, m}\; \mathbb{E}_{x}\Big[ \big\| R_\theta\big(\mathcal{P}_m \mathcal{F} x,\, m\big) - x \big\|_1 \Big], $$
where $x$ is a fully sampled reference scan, $\mathcal{F}$ is the Fourier transform, $\mathcal{P}_m$ keeps only the k-space lines selected by the learnable mask $m$, and $R_\theta$ is an unrolled reconstruction network whose iterations alternate learned regularization blocks with hard data-consistency layers that force agreement with every actually acquired sample — the design that dominated the cardiac reconstruction challenge (CMRxRecon results). The bowel-specific novelty is in what the mask is allowed to learn from: we train on paired scans of dysbiotic and control animals, so the learned pattern can preferentially sample the k-space regions that carry wall-contrast information and be robust to the susceptibility and motion conditions of the gas-filled dysbiotic abdomen, a regime the existing scan-adaptive methods (SUNO; T1-PILOT) never evaluated. Because deep reconstructions at high acceleration are known to over-smooth or invent anatomy — the 2020 fastMRI challenge documented networks hallucinating vessels out of metal clips at $8\times$ (Results of the 2020 fastMRI Challenge for Machine Learning MR Image Reconstruction) — every map ships with a per-voxel uncertainty estimate, and the pipeline is evaluated not on image prettiness but on relaxation-time accuracy and on a synthetic-defect preservation test described below.
Third, biological validation in the two mouse models that anchor the literature. The meropenem arm follows the Ribeiro design: MDR E. coli infection in BALB/c mice, five days of meropenem, with scans at baseline, end of treatment, day 14, and day 30. The diarrhea arm follows the ELF051 protocol: a seven-day amoxicillin/clindamycin/streptomycin cocktail in C57BL/6 mice, with a parallel probiotic-intervention group. Because histology is terminal, each scan time point is anchored by a sacrificed subcohort whose crypt depth, goblet-cell counts, TLR4/MyD88/NF-$\kappa$B pathway protein expression, and cecal short-chain-fatty-acid concentrations (the ELF051 study reports 1559.58 µg/g in untreated antibiotic-associated diarrhea versus 1835.74 µg/g under probiotic rescue) provide ground truth. Fecal 16S sequencing runs alongside as a covariate, not a competitor. All raw k-space, masks, and parameter maps will be released — the file identifies the near-total absence of gastrointestinal k-space benchmarks as an open problem, and this dataset would be the first.
Baselines
The naive reference point is low-pass k-space filtering of the undersampled data — keep the center of k-space, discard the rest, accept the blur. On the fastMRI knee dataset at $80\times$ acceleration, Huijben et al. report a peak signal-to-noise ratio of 35.8 dB for this baseline, versus 27.7 dB for uniform random sampling and 36.2 dB for their learned approach; we expect the same ordering on bowel data, and the low-pass map is what a radiologist would get today from any scanner asked to rush. It is the strongest trivial competitor because it is completely artifact-deterministic, but its blurring is precisely what erases a sub-millimeter wall, so beating it means demonstrating higher relaxation-time accuracy against the spin-echo reference at matched scan time, not just higher structural similarity.
The strongest existing methodological competitor is the pipeline of Stelter et al.: a physics-model-based sequence with total-variation-regularized subspace reconstruction, water–fat separation, and field-inhomogeneity-corrected dictionary matching, which achieved mean water $T_1$ accuracy of (12 ± 22) ms and water $T_2$ accuracy of (−1.4 ± 3.1) ms against spin-echo references in a phantom, at a 3-mm whole-abdomen scan of 5 minutes 23 seconds. We will run their reconstruction unchanged on our murine data at matched scan duration and matched total sampling budget; this is the comparison that matters because our endpoint is parameter accuracy in milliseconds, not image quality. Alongside it we run the strongest learned reconstruction from the challenge literature — an end-to-end variational network of the kind that won the CMRxRecon cine task with a structural similarity of 0.990 ± 0.002 (CMRxRecon results) — but with a fixed heuristic undersampling pattern, to isolate what the sampling learning contributes. All methods receive the same training scans, the same GPU compute, and the same hyperparameter search budget, and are scored on the same held-out animals.
The key ablation swaps our learned sampling mask for a fixed uniform-random mask of identical sampling fraction while holding the reconstruction network fixed, which isolates the single ingredient this proposal adds over both competitors. Huijben et al. showed an 8.5 dB gap between learned and uniform-random sampling on knee data at extreme acceleration; if no analogous gap (in relaxation-time error, not just PSNR) appears on bowel data, the learned-sampling half of the claim is dead and the proposal reduces to a sequence-transfer study — we would report it that way. A second ablation removes the data-consistency layers, testing whether strict fidelity to acquired samples is what keeps genuine mucosal defects from being smoothed away.
Why not just…?
Why not just track dysbiosis with fecal 16S sequencing and metabolomics? Because fecal content measures the lumen, not the wall, and the wall is where the clinical risk lives. In the meropenem model, fecal sequencing correctly flagged the persistent dysbiosis but could not say whether the barrier was intact or leaking, while systemic leukocyte counts — the marker a blood draw would give — had fully normalized by day 30 (Ribeiro et al.). Fecal material also averages the entire digestive tract into one sample, erasing the segmental variation the ELF051 study demonstrates between cecum and colon (ELF051 study), and sequencing turnaround is days, not minutes. That said, this alternative is a real complement rather than a defeated rival: fecal profiling stays in our study as a covariate, and the decisive statistical test is whether segmental wall $T_2$ predicts histological outcome beyond what fecal composition already predicts. If it does not add predictive power, the imaging is redundant and we will say so.
Why not just use classical compressed sensing and skip the deep networks? Classical total-variation regularization is mathematically understood and cannot hallucinate, which is a genuine advantage given that the 2020 fastMRI challenge documented deep networks synthesizing anatomy at high acceleration (Muckley et al.). But the fidelity cost is large and well quantified: even a basic U-Net beat the total-variation baseline by 40–50% relative improvement in normalized mean squared error on fastMRI (Zbontar et al.), and total variation's staircasing artifact turns thin structures into flat "waxy" slabs — exactly the failure mode that would erase a sub-millimeter mucosal layer. We take the hallucination concern seriously rather than arguing around it: the synthetic-defect experiment below directly measures, for each method, the rate at which real inserted defects are smoothed away versus false defects invented, so if classical compressed sensing wins on that clinically meaningful trade-off at modest acceleration, the data will show it.
Why not just run a standard fully sampled scan with sedation and antispasmodic drugs? Full sampling takes 30–60 minutes (Puy et al.), during which peristalsis corrupts k-space in ways no retrospective correction can undo, and repeated general anesthesia across a longitudinal four-timepoint study is both a welfare burden and, in pediatric and fragile populations, an independent clinical risk (Qiyang Zhang et al.). Worse for this specific application, motility-suppressing drugs alter gut transit, mucosal blood flow, and fluid balance — they pharmacologically perturb the barrier physiology we are trying to measure. The existence proof that none of this is necessary is Stelter et al.'s free-breathing three-to-five-minute abdominal protocol (Stelter et al.); our question is whether that feasibility extends from the liver to the moving, gas-filled gut, and under how much acceleration.
Decisive Checks
Four experiments decide the proposal. The first is calibration: on a water–fat phantom and on a small fully scanned animal cohort, the accelerated pipeline must recover spin-echo reference $T_1$/$T_2$ within an error envelope comparable to Stelter's reported (12 ± 22) ms and (−1.4 ± 3.1) ms at a scan time at or under their 5-minute-23-second clinical figure, at acceleration factors from $4\times$ to $8\times$; the last fully sampled scans serve as reconstruction ground truth scored by structural similarity and normalized mean squared error against the fully sampled image, the standard fastMRI metrics (Zbontar et al.). The second is the hallucination test motivated by Gap 4: we digitally insert thin, wall-scale synthetic defects into fully sampled bowel scans, retrospectively undersample, and measure each method's defect-preservation and false-positive rates — this is where the data-consistency ablation and the uncertainty maps earn or lose their place. The third is the biological endpoint: at day 30 in the meropenem arm, segmental wall $T_2$ (with $T_1$ as secondary) must discriminate post-antibiotic mice from uninfected controls above a pre-registered area-under-the-curve threshold, and must correlate with terminal crypt depth, goblet-cell counts, and pathway protein expression in the anchored subcohort — this is the direct test of whether relaxometry sees the scar that clinical markers miss. The fourth is the intervention arm: the pipeline must detect the recovery direction in the probiotic-treated diarrhea group consistent with the known metabolite rescue from 1559.58 to 1835.74 µg/g and the 17–22% reductions in TLR4, MyD88, and NF-$\kappa$B expression (ELF051 study) — if the maps move with treatment, longitudinal non-invasive monitoring of gut-barrier recovery becomes a concrete, reproducible capability rather than a claim about one timepoint. No reported result found on Lacuna for any prior quantitative MRI measurement of the bowel wall in a dysbiosis model, which is precisely the space this work opens.
References
- Absence of gut microbiota restoration following meropenem treatment in a systemic infection model
- Spread Spectrum Magnetic Resonance Imaging
- fastMRI: An Open Dataset and Benchmarks for Accelerated MRI.
- The state-of-the-art in cardiac MRI reconstruction: Results of the CMRxRecon challenge in MICCAI 2023
- Abdominal simultaneous 3D water T 1 and T 2 mapping using a free‐breathing Cartesian acquisition with spiral profile ordering
- Lactiplantibacillus plantarum ELF051 Alleviates Antibiotic-Associated Diarrhea by Regulating Intestinal Inflammation and Gut Microbiota
- Scan-Adaptive MRI Undersampling Using Neighbor-Based Optimization (SUNO)
- T1-PILOT: Physics-Informed Learned Optimized Trajectories for T1 Mapping Acceleration
- Results of the 2020 fastMRI Challenge for Machine Learning MR Image Reconstruction
- Learning Sampling and Model-Based Signal Recovery for Compressed Sensing MRI
- Reducing pediatric total-body PET/CT imaging scan time with multimodal artificial intelligence technology