Surgeons repairing a severed nerve across a long gap usually take a healthy nerve from elsewhere in the patient's body, most often the sural nerve in the leg. This "autograft" leaves permanent numbness, risks a painful neuroma at the harvest site, and is in limited supply. Engineered tubes that guide regrowing nerve fibers, called nerve guidance conduits, now match autografts in rat experiments. Surgeons still hesitate to use them first. If a conduit fails, nobody finds out for months, and by then the target muscle may have wasted beyond rescue. This proposal asks for a fast, cheap, repeatable bedside test that says within the first few weeks whether a conduit is on track. With such a test, surgeons could implant the conduit first and harvest a donor nerve only for the minority whose conduits are failing.
The core idea fits in two sentences. We image implanted conduits in rats every week with high-frequency ultrasound and train a small model to predict, from the first four weeks of images, the functional outcome measured at twelve weeks. The test that matters is whether this prediction works within a single conduit type. That would show it reads the tissue rather than simply recognizing which material was implanted, and that it beats simpler early signals such as hand-measured nerve thickness or early electrophysiology.
Core Claim
Current conduit studies judge success only at the end, using post-mortem histology and functional tests at 12 weeks. Examples are the Matrilin-2/lysine-chitosan scaffolds of Neill Y. Li et al. and the conductive stem-cell conduits of Hong Cheng et al.. Clinically, the information arrives too late to act on, so conduits stay a risky first choice and autograft harvest stays the default.
We claim that early ultrasound changes carry enough information to sort conduits into "likely to fail" and "likely to succeed" by week four. The relevant changes are whether a tissue cable fills the conduit lumen, its echo texture, swelling of the proximal stump, and the size of the distal nerve. Failure is defined, before any data are collected, as 12-week compound muscle action potential (the electrical response of the target muscle to nerve stimulation) amplitude below a fixed fraction of the healthy side.
The model is borrowed from ultrasound nerve-segmentation work. What is new is the question it answers: predicting regeneration outcome in an implanted scaffold, with the clinical aim of a conduit-first, harvest-on-failure pathway.
The model is not expected to see individual axons; ultrasound cannot resolve them. The bet is that coarse tissue changes at the millimeter scale are predictive enough. If that bet is wrong, the proposal fails cleanly.
Relation to Prior Work
The nerve-ultrasound literature supplies the method but studies a different problem. Region-aware Global Context Modeling for Automatic Nerve Segmentation from Ultrasound Images segments healthy nerves for needle guidance. It replaces pixel-level attention with pooling over 3×3 to 11×11 regions so that speckle noise is not amplified, and it reaches a Dice of 74.23%. Investigation of Appropriate Scaling of Networks and Images for Convolutional Neural Network-Based Nerve Detection in Ultrasound-Guided Nerve Blocks finds that downscaled 384×384 inputs with a large Scaled-YOLOv4-P7 detector work best, because downscaling smooths speckle. Temporal Monte Carlo Dropout estimates per-frame uncertainty cheaply by spreading dropout samples across consecutive video frames, at one eighth of the compute, but it assumes a stationary probe. All three study intact nerves at a single moment. None follows a repaired nerve over time or predicts an outcome. We reuse their segmentation backbones and their uncertainty trick and point them at a longitudinal prediction task.
The scaffold work motivates the proposal and also creates its main confound. Li et al. and Cheng et al. show that well-designed conduits can rival autografts, while untreated collagen conduits do much worse. That gap between material types is exactly why a naive predictor could look good just by recognizing the material. Machine intelligence for nerve conduit design and production is the closest computational bridge. It reviews 7T MRI fascicle segmentation for designing patient-specific conduits and reinforcement-learning predictors of graft success. All of that is preoperative and depends on centralized high-field scanners. It notes "small data" as a central bottleneck, which constrains our design toward small models and engineered features. Label Uncertainty for Ultrasound Segmentation supplies a training detail: weighting annotations by clinician confidence improved downstream clinical prediction. This matters here because the edge between ingrowing tissue and a degrading hydrogel wall is genuinely ambiguous.
The diagnostic logic is taken from Osamu Nomura et al.. Their meta-analysis of the 0/1-hour troponin algorithm shows that a fast test is valuable when it is tuned for safe rule-out (pooled sensitivity 99.3%), not for overall accuracy. Our operating point is chosen the same way. Missing a failing conduit costs the patient the muscle. A false alarm costs an early re-exploration, which is the procedure surgeons do today anyway.
Research Program
Animal cohort. We build a longitudinal rat sciatic-nerve dataset, because none exists. Gap injuries are bridged with reverse autografts, untreated hollow collagen conduits, and a bioactive conduit following the Matrilin-2/lysine-chitosan design of Li et al. We use two gap lengths so that outcomes spread widely within each material, not only between materials. The target is roughly 15 animals per arm, around 90 in total.
Imaging. Each animal is scanned under brief anesthesia at weeks 1, 2, 3, 4, 6, 8, and 12 with a high-frequency linear transducer. The probe is mounted on a motorized stage that sweeps along the repair at constant speed. This deliberately sidesteps the stationary-probe limitation of the Temporal Monte Carlo Dropout work while keeping the sweep reproducible. Endpoint labels are the 12-week muscle response on the treated side relative to the healthy side, walking-track force, target muscle weight, and distal axon counts. These are the same endpoints Li et al. report.
Model. The analysis runs in two stages, a choice made because the small dataset rules out an end-to-end video model.
In the first stage, a segmentation network is pretrained on the Kaggle ultrasound nerve benchmark of 11,143 images. It uses the ResNet-34 U-Net with region-aware pooling from Wu et al. It is then fine-tuned on a few hundred frames from our sweeps annotated by two sonographers, who mark stumps, conduit wall, lumen contents, and any neuroma bulb, with confidence-weighted loss. Uncertainty comes from Temporal Monte Carlo Dropout across sweep frames.
In the second stage, segmentation is turned into a short list of interpretable trajectories: lumen fill fraction, mean and texture statistics of tissue echo intensity inside the lumen, proximal stump swelling, distal nerve cross-sectional area, and conduit wall thickness as a proxy for resorption. A sparse logistic model or small Gaussian-process classifier then predicts failure from weeks 1–4. Animals whose segmentation uncertainty exceeds a threshold are flagged "indeterminate" rather than forced into a call, mirroring the troponin algorithm's observe zone.
Resolution. The resolution tension raised by the downscaling result is tested directly. We run the pipeline at 384×384 and at native resolution. Downscaling may help find the nerve while erasing lumen texture that may carry the outcome signal.
Baselines
The naive reference point is a material-and-gap-only predictor: a logistic model that knows which conduit and gap length were used and nothing else. Because material type strongly predicts outcome in Li et al., this baseline may already look respectable on pooled data. It is also the main threat. Any ultrasound model must beat it, and must discriminate within each conduit arm, where this baseline has no information at all.
The strongest simple competitor is hand-measured ultrasound biometrics. A blinded sonographer measures distal nerve cross-sectional area and stump diameter at weeks 1–4 using standard calipers, and the same logistic model is fit. It sees the same images, the same timepoints, and the same cross-validation folds. Our pipeline beats it only if within-arm AUC is higher with a bootstrap confidence interval excluding zero difference. We will also report the hand-measurement time per animal.
The strongest existing automated method is off-the-shelf nerve segmentation features, in two versions. One is the Wu et al. architecture trained only on Kaggle and applied frozen. The other is Scaled-YOLOv4-P7 at 384×384, the configuration that the scaling study reports at F1 94.7% on public nerve data and 80.8% on brachial plexus data. Both get the same downstream feature extraction and classifier. For reference, Wu et al. report Dice 74.23% on Kaggle. For our actual task, predicting regeneration outcome from ultrasound, no reported result found on Lacuna, so the baselines must be run by us.
The early electrophysiology baseline measures the muscle response at week 4 and week 6 under the same anesthesia session.
Ablations isolate the key ingredients. We drop the longitudinal trajectory and use only the week-4 scan. We remove confidence weighting. We run at 384×384 versus native resolution. We remove the abstention rule, so that we can see how much apparent accuracy depends on discarding hard cases.
All models share the same nested leave-animals-out cross-validation and an equal hyperparameter search budget of a fixed number of configurations.
Why not just…?
Why not just harvest the sural nerve, as surgeons do now? Harvest is reliable but carries permanent donor-site numbness and neuroma risk, and supply runs out in multi-nerve or branching injuries (Lingchi Kong et al.). Engineered conduits already match reverse autografts in rats (Neill Y. Li et al.). What is missing is a way to catch the cases where they don't. This proposal does not argue that autografts should be abandoned. It argues that an early test would let surgeons reserve them for the cases that need them.
Why not just have a sonographer measure nerve thickness by hand? This is a real threat. If distal cross-sectional area at week 4 predicts failure as well as the learned pipeline, the learned pipeline adds only speed and consistency, and we will say so. The within-arm comparison against hand biometrics in the Baselines section settles it. Our prior is that lumen fill and texture carry information that a single diameter misses, but this is an empirical question.
Why not just wait for early electrophysiology? Muscle responses depend on axons actually reaching the muscle. Before reinnervation they are absent in both future successes and future failures, which is why acute electrodiagnostics cannot tell conduction block from disruption. If the week-6 muscle response turns out to be predictive, the useful comparison becomes whether ultrasound gives a comparable answer two weeks earlier. Our timepoint design measures exactly that.
Why not just use MRI? The 7T fascicle-mapping pipeline reviewed in Machine intelligence for nerve conduit design and production is preoperative and needs centralized scanners. Weekly high-field imaging is not a realistic clinical follow-up, and a bedside test is the point of this work.
Proposed Experiments and Decisive Checks
Primary endpoint. The primary endpoint is week-4 AUC for predicting 12-week failure, computed within each conduit arm and then averaged. This within-arm check is decisive. If the ultrasound model is no better than the material-and-gap predictor inside arms, the project's central claim is false.
Clinical operating point. We report sensitivity for failure at the threshold that keeps the false-alarm rate below a pre-set ceiling, together with the fraction of animals flagged indeterminate. A model that is accurate only after abstaining on half the cohort is not useful.
Generalization across materials. We train on autograft and collagen arms and test on the bioactive-conduit arm. This asks whether the features capture tissue regeneration rather than material-specific echo signatures, which matters because degrading hydrogels change echogenicity on their own.
Agreement with histology. We test whether week-12 lumen fill and texture features correlate with distal axon counts, so that the image features have a biological interpretation.
Statistical power. With roughly 90 animals, the confidence interval on AUC will be about ±0.1. The proposal is therefore powered to detect a clear signal, not small improvements. Effect sizes and bootstrap intervals are pre-registered.
References
- Matrilin-2 with a K-Chitosan Scaffold Enhances Functional Recovery and Nerve Regeneration in a Segmental Rat Sciatic Nerve Injury Model
- Conductive Nerve Guidance Conduits Loaded With Adipose Mesenchymal Stem Cells for Peripheral Nerve Regeneration
- Region-aware Global Context Modeling for Automatic Nerve Segmentation from Ultrasound Images
- Investigation of Appropriate Scaling of Networks and Images for Convolutional Neural Network-Based Nerve Detection in Ultrasound-Guided Nerve Blocks
- Temporal Monte Carlo Dropout for Robust Uncertainty Quantification: Application to Point-of-Care Ultrasound-guided Nerve Blocks
- Machine intelligence for nerve conduit design and production
- Label Uncertainty for Ultrasound Segmentation
- Performance of the 0-Hour/1-Hour Algorithm for Diagnosing Myocardial Infarction in Patients With Chest Pain in the Emergency Department ― A Systematic Review and Meta-Analysis ―
- Multilevel neurium-mimetic individualized graft via additive manufacturing for efficient tissue repair