Epithelial tissues — the gut lining, the blood-brain barrier, the lung's air surface — are sealed by tight junctions: interlocking protein strands (occludin, claudins, and the scaffolding protein ZO-1) that zip adjacent cells together and control what leaks between them. When inflammation or diet-driven stress unzips these junctions, bacterial toxins enter the bloodstream and drive systemic disease, and a large preclinical literature shows that restoring junction protein expression reseals the barrier: nuciferine treatment in high-fat-diet rats cut serum triglycerides by 46% while raising ZO-1 and occludin (Nuciferine modulates the gut microbiota), and Gpr174 knockout protected mice from chemically induced colitis while upregulating the same proteins (Gpr174 Knockout Alleviates DSS-Induced Colitis). Yet these successes collapse in human studies, where cohorts are small, events are rare, and endpoints are noisy scalar surrogates — exactly the conditions under which statistical power, the probability of detecting a true effect, disintegrates (Peter B. Gilbert et al.; Zhe Huang et al. documented a nearly five-month numerical survival benefit that failed significance for lack of power).
This proposal argues that the endpoint itself is the bottleneck, and that architecture search is the right tool to fix it. The idea: run direct large-scale architecture search — searching network designs on the target task itself, using memory-bounded path sampling rather than proxy datasets — directly on confocal immunofluorescence images of tight junctions in small biopsy cohorts, with a topology-aware objective that rewards junction continuity rather than pixel volume, and distill the network's output into a continuous Barrier Integrity Index. The hypothesis is that this index raises the standardized effect size (signal divided by noise) per patient, so that barrier-recovery trials need far fewer subjects to reach significance. In one sentence: search the segmentation architecture directly on tight-junction images so that every pixel of an expensive biopsy is converted into a low-variance continuous measurement instead of a coarse, noisy scalar.
Core Claim
Two failure modes currently compound each other. First, the biology has a measurement problem: quantitative protein upregulation on a Western blot does not guarantee a functional seal, because junction proteins can accumulate in disorganized cytosolic pools without polymerizing into an intercellular zipper (Nanotopography enhances dynamic remodeling of tight junction proteins); and bulk functional readouts like serum endotoxin or whole-tissue electrical resistance average away exactly the localized structural remodeling that early therapy produces. Second, the vision tooling has a geometry problem: cell segmentation models and benchmarks are built for cell interiors and membranes (The Multi-modality Cell Segmentation Challenge; FCIS), not for one-pixel-wide reticular junction meshes, where breaking a single strand opens the barrier but costs almost nothing under a Dice or cross-entropy loss.
The core claim is that both problems are addressable in one integrated system, and that the payoff is measured in patients per detected effect, not in segmentation Dice. For a two-arm trial, the required per-arm sample size is roughly
$$n \approx \frac{2\,(z_{1-\alpha/2} + z_{1-\beta})^2\, \sigma^2}{\delta^2},$$
where $\delta$ is the true treatment effect on the endpoint and $\sigma$ its within-arm spread; the $1/\delta^2$ form is the standard approximation behind power analysis. An image-derived continuous index attacks both terms: it preserves structural effect magnitude (detecting focal strand repair that bulk assays average out) while shrinking measurement variance relative to systemic biomarkers that carry hepatic, dietary, and metabolic confounding. Halving $\sigma/\delta$ quarters the required cohort — the difference between an underpowered pilot and an interpretable trial.
Research Program
The program has three components. First, a benchmark: we will assemble a tight-junction imaging corpus from the two animal models documented in the literature, DSS colitis and high-fat-diet rats (Wei et al., whose RNA-seq is deposited under accession PRJNA819571; Wang et al.), plus LPS-challenged epithelial monolayers as in the in vitro work of Hengli Guo et al.. Each confocal ZO-1/occludin field comes with paired functional ground truth (TEER and dextran flux where monolayers permit, serum endotoxin and histology scores where tissue permits) and human-annotated junction strands. Critically, labels are graded on a continuous severity scale rather than binary intact/disrupted, following the finding of Kang et al. that hard binary labels on heterogeneous small cohorts destroy statistical sensitivity — their soft-target method in dementia speech detection lifted balanced accuracy to 0.8608 on ADReSS and recall from 0.4615 to 0.5881 on CU-MARVEL, so we adapt their ensemble distillation to imaging rather than reinventing it.
Second, the search itself. We adopt the machinery of proxyless search — path binarization bounding memory to a single network regardless of search-space size (ProxylessNAS) and quarter-channel sampling with edge normalization for stable gradients (PC-DARTS) — but over a densely connected multi-scale space in the spirit of DCNAS, which searched directly on high-resolution Cityscapes segmentation in 5.6 GPU days and confirmed search-phase/final-performance rank correlation of Pearson $\rho = 0.73$. The search objective is where we depart from all of these: the loss combines a strand-continuity term (a connectivity-sensitive penalty on the junction network's skeleton, so that breaking a seal costs disproportionately) with a distillation term tying the network's pooled features to the paired functional measurement. We deliberately avoid the ImageNet-proxy route because the proxy gap is documented to be worst exactly where geometry is unusual — dense, high-aspect-ratio structures at high resolution (DCNAS).
Third, the power analysis as a first-class deliverable. Using uncertainty quantification by hypothesis testing (HashUQ), we will estimate, per endpoint, the detectable effect size at fixed $N$ and the required $N$ at fixed power, comparing the Barrier Integrity Index against serum endotoxin, TEER, and Western blot densitometry on the same animals. The preclinical models give large effect sizes ($d > 2$ under massive chemical insult), so we will additionally run a powered-down analysis: subsampling to synthetic small-cohort regimes that mimic the human trials in the literature, where rare events and heterogeneity already defeat scalar endpoints (Gilbert et al.; Nseir et al.).
Baselines
The do-nothing reference point is the field's current practice: bulk endpoints alone. Serum endotoxin, cytokines, and Western blot ZO-1/occludin densitometry are exactly how the motivating studies quantified barrier rescue (Wang et al.; Wei et al.), and no reported numerical fold-change for junction protein upregulation is available on Lacuna for either — which itself illustrates the reporting problem. We will compute the detectable effect size of these scalar endpoints on our paired data; beating the reference means achieving significance at a cohort size where these endpoints do not, and, more basically, resolving the upregulation-versus-assembly contradiction that bulk densitometry cannot (Nanotopography).
The strongest segmentation competitor is nnWNet (nnWNet: Rethinking the Use of Transformers in Biomedical Image Segmentation), which reported statistically significant improvement in Jaccard, Dice, and 95th-percentile Hausdorff distance in 87.5% of cases across eight standardized biomedical datasets, with 7M parameters versus 92M for the nnUNet baseline — a 13x reduction that matters greatly on small biopsy cohorts. Its official codebase is public, so we run it with identical training data, identical tuning budget (matched GPU-hours to our search-plus-training total, mirroring the budget-reporting convention of PC-DARTS at 0.1 GPU-days for search), and identical evaluation. Beating it requires two things: better junction-continuity and topology metrics at equal Dice (a fixed nnWNet trained with a volumetric loss should systematically under-seal thin strands), and a larger power gain per patient when its outputs are distilled into a continuous index — if nnWNet plus our soft-score head matches our searched network on power, the search component is not earning its cost, and we will say so. A second fixed-architecture baseline is the SegFormer-based hierarchical encoder with cell memory replay that won the Multi-modality Cell Segmentation Challenge with a median F1 of 89.7% (Ma et al.), run under the same protocol, because it represents the strongest transferable cell-boundary model in the corpus.
The key ablation isolates the proposal's two ingredients separately. The discovered architecture is retrained with a plain Dice loss and binary labels (no continuity objective, no soft distillation) — this separates architecture from objective; and the best fixed baseline is given our continuity objective and soft-score head — this separates objective from architecture. The claim is vindicated only if both ingredients contribute, and that the combination beats each single-ingredient variant on the power metric.
Why not just…?
Why not just use off-the-shelf nnUNet or SegFormer without architecture search? This is the most serious objection and we treat it as such rather than arguing around it. nnWNet's 7M-parameter design was hand-built for general biomedical boundaries (nnWNet); it may transfer well to junction strands. The counter-evidence is that fixed multi-scale topologies were exactly what direct dense search beat on unusual boundary geometry — DCNAS surpassed DeepLabv3+ on PASCAL VOC 2012 with 86.9% mIoU precisely because cross-layer connectivity was searched, not fixed (DCNAS) — but that evidence is from natural images, not junctions. Our ablation (best fixed architecture plus our objective) settles this empirically: if search adds nothing beyond the objective, we will report the fixed model with the continuity loss as the recommended method and the search as negative result.
Why not just measure TEER or serum endotoxin and skip imaging entirely? These assays measure flux directly and are cheaper per sample. But they are the endpoints that already fail in small human cohorts: systemic biomarkers carry large between-subject variance from hepatic and dietary confounders (Wang et al. for the biomarker chain; Nseir et al. for variance suppressing power in sepsis cohorts), and whole-tissue scalars are structurally blind to focal strand repair versus cytosolic protein pooling (Nanotopography). Our decisive experiment is the head-to-head power curve on paired data: if serum endotoxin detects the same effect at the same $N$ as the image index, imaging loses and we will say so. We also note that TEER is unavailable on human endoscopic biopsies in practice, while immunofluorescence is routine.
Why not just apply permutation tests or bootstrapping to the existing scalar endpoints? Classical small-sample corrections cannot recover power that the measurement already discarded; they control error rates on whatever signal survives. The degrees-of-freedom and heterogeneity failures documented in small clinical subgroups (Yuan Yang et al.; Amir Kaywan Aftahy et al.) apply to the endpoints, not the test; and Kang et al. showed standard ensembles and classical baselines underperform continuous soft scoring under within-class heterogeneity (recall 0.4615 versus 0.5881). We will nevertheless run exact permutation tests on our index to confirm the gain is from the endpoint, not from a liberal test.
Why not just fine-tune a pathology foundation model? The available healthcare foundation models, catalogued in Awesome-Foundation-Models-for-Advancing-Healthcare and exemplified by the spatial-transcriptomics model SToFM, are trained on H&E tissue-scale morphology, roughly two orders of magnitude coarser than sub-micron immunofluorescent junction strands, and the Multi-modality Challenge showed even strong vision models transfer poorly across microscopy modalities without modality-specific adaptation (Ma et al.). We will still fine-tune the strongest available pretrained encoder as an additional baseline under matched budget, since a reviewer is entitled to that comparison; if it wins, the field should use it and the search-space contribution is falsified.
Decisive Checks
Three results would make or break the proposal. First, the power curve: on paired animal data subsampled to synthetic small cohorts, the Barrier Integrity Index must achieve 80% power at a materially smaller $n$ than serum endotoxin or densitometry, with the gap explained by effect-size preservation (correlation of the index with paired TEER/dextran-flux ground truth comparable to or better than the clinical correlations SoTD achieved, $r = 0.7375$ against MMSE in Kang et al.). Second, the topology check: the searched-with-continuity-loss model must beat the same model with Dice loss, and fixed baselines, specifically on junction-seal metrics — strand breakage rate and seal connectivity — where a single broken pixel is functionally catastrophic, not merely on Dice where all models may tie. Third, the search-validity check: rank correlation between search-phase and final-model performance should meet or exceed the $\rho = 0.73$ benchmark DCNAS established for direct dense search (DCNAS), at a search cost reportable in GPU-days against DCNAS's 5.6 and ProxylessNAS's 200 GPU hours (ProxylessNAS). If the first check fails, the premise is wrong; if the second fails, the objective is wrong; if the third fails, direct search does not transfer to this regime — and each failure is independently reportable.
References
- Nuciferine modulates the gut microbiota and prevents obesity in high-fat diet-fed rats
- Gpr174 Knockout Alleviates DSS-Induced Colitis via Regulating the Immune Function of Dendritic Cells
- HAI and NAI titer correlates of inactivated and live attenuated influenza vaccine efficacy
- Benefit assessment of extended dosing in cancer patients after their withdrawal from clinical trials
- Nanotopography enhances dynamic remodeling of tight junction proteins through cytosolic complexes
- The Multi-modality Cell Segmentation Challenge: Towards Universal Solutions
- Kuwanon G Preserves LPS-Induced Disruption of Gut Epithelial Barrier In Vitro
- On the Within-class Variation Issue in Alzheimer's Disease Detection
- ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware
- PC-DARTS: Partial Channel Connections for Memory-Efficient Architecture Search
- DCNAS: Densely Connected Neural Architecture Search for Semantic Image Segmentation
- nnWNet: Rethinking the Use of Transformers in Biomedical Image Segmentation and Calling for a Unified Evaluation Benchmark
- Correlation of 5-HTT, BDNF and NPSR1 gene polymorphisms with anxiety and depression in asthmatic patients
- Intraventricular neuroepithelial tumors: surgical outcome, technical considerations and review of literature