Does Lithotripsy's Tissue Selectivity Survive the Spinal Canal? A Transportable Wave-Surrogate Study

Intravascular lithotripsy cracks calcium inside arteries by firing small acoustic shockwaves from a balloon. It is safe mainly for one physical reason. Soft vessel wall transmits sound almost like water, so the wave passes through it harmlessly. Calcium has very different acoustic properties, so the wave reflects off it and concentrates stress there. Spine surgeons working through narrow endoscopic corridors face a similar problem: calcified disc fragments and osteophytes (bony spurs) sit millimetres from the dura (the membrane around the nerves). Their tools are burrs and rongeurs, which can tear the dura or overheat nerves. The obvious idea is to bring lithotripsy into the spinal canal. Before anyone builds a device, though, one question has to be answered. Does the impedance-mismatch selectivity that protects an artery still protect the dura? A calcified disc fragment is backed by vertebral bone, surrounded by fat and cerebrospinal fluid, and confined by a bony canal that can trap and refocus reflected waves.

We propose to answer that question computationally, at the level of individual patients, and to make the answer trustworthy across sites. The instrument is a neural-operator wave surrogate: a network that learns to map a tissue map directly to the resulting pressure field, in milliseconds instead of hours. The surrogate is needed for uncertainty quantification over unknown tissue properties, not for single forward runs. Its credibility is tested the way clinical prediction models should be tested and rarely are. The model is frozen, then evaluated on anatomy from a different domain and a different center, with a pre-specified tolerance. Geophysics supplies both the method and the warning. Neural operators emulate wave propagation well, but Velocity continuation with Fourier neural operators for accelerated uncertainty quantification describes its own operator as survey-specific.

Core Claim

The claim is that whether lithotripsy is selective in the spinal canal can be stated as a per-patient, per-emitter-position quantity. That quantity can be computed with honest uncertainty from routine CT. We will define the selectivity ratio

$$S = \frac{\max_{x \in \Omega_{\text{calc}}} \sigma_1(x)}{\max_{x \in \Omega_{\text{dura}}} \sigma_1(x)},$$

where $\sigma_1(x)$ is the peak maximum principal stress at location $x$ over the shock pulse. $\Omega_{\text{calc}}$ is the calcified target region and $\Omega_{\text{dura}}$ is the dural sac boundary.

In arteries the analogous ratio is favourable by design. We hypothesize that in the spine $S$ varies widely across three factors: the patient's anatomy, the position of the emitter in the corridor, and the bone mineral density behind the fragment. We further hypothesize that a meaningful fraction of realistic configurations are unsafe even though each layer, taken on its own, has a favourable impedance contrast. If $S$ is uniformly high, the device case is strong. If it is low and unpredictable, the idea should be dropped before animal work. Either outcome is useful, and neither is in the literature.

How the Pieces Fit

Lithotripsy supplies the mechanism and the target for validation. Impact of lesion preparation-induced calcified plaque defects in vascular intervention for atherosclerotic disease: in silico assessment built 22 parametric finite-element models calibrated to optical coherence tomography (OCT, an intravascular imaging method) from 262 Disrupt CAD-III patients. It showed that full-thickness "through-cracks" act as hinges, giving up to 51% relative lumen gain versus 16–25% for partial cracks. It also showed that crack length matters more than crack width.

That study inserted cracks into the mesh by hand and never simulated the shockwave itself. We add the missing wave step and ask a further question: in the spine, does the stress field produce a through-fracture of the fragment while leaving the dura below a damage threshold?

Structural spinal preservation is the application setting. Endoscopic discectomy can let compensatory scoliosis resolve in over 90% of patients without fusion (Percutaneous endoscopic lumbar discectomy for single and double segment lumbar disc herniation with sciatic scoliosis in adults: a retrospective study). Corridor approaches such as oblique lumbar interbody fusion (OLIF) reduce blood loss and operative time compared with transforaminal fusion (Meta-analysis of minimally invasive transforaminal lumbar interbody fusion versus oblique lumbar interbody fusion for treating lumbar degenerative diseases). Both depend on avoiding bone resection, and both stall at calcified pathology. Neither study involves any wave physics; we supply it.

Patient-specific geometry and material properties come from methods that map CT intensity to stiffness. Patient-Specific Finite Element Modeling of the Whole Lumbar Spine Using Clinical Routine Multi-Detector Computed Tomography (MDCT) Data—A Pilot Study showed that routine CT predicts vertebral failure load ($R^2 = 0.8$ against bone density). Its models were static, homogeneous, and built from 12 subjects at one center, which is exactly the situation external validation is meant to expose.

Geological generalizability supplies both the surrogate method and a cautionary prior. Neural operators handle wave propagation in layered media, but they can become survey-specific. Simpler learners show the same collapse on a new well: a decision tree fell from $R = 0.99$ in training to $0.54$ externally, while a random forest held $0.71$ (Prediction of Young’s Modulus for Reservoir Rocks using Machine Learning).

External cohort validation supplies the protocol. A model that scored a mortality AUC of about 0.90 internally dropped to 0.74 on UK Biobank (Developing and validating COVID-19 adverse outcome risk prediction models from a bi-national European cohort of 5594 patients). Only 77 of 251 prostate radiomics papers tested on independent cohorts at all (Radiomics in prostate cancer imaging for a personalized treatment approach - current aspects of methodology and a systematic review on validated studies). These papers motivate the evaluation design, but none of them involves physics.

Research Program

First, we build a reference solver pipeline. Spine geometries are segmented from CT, and each tissue class is assigned acoustic and elastic properties: calcified fragment, cortical and cancellous bone, disc, ligamentum flavum, epidural fat, cerebrospinal fluid, dura, and the saline in the endoscopic working channel. Bone properties come from CT intensity, following the MDCT pilot approach. A full elastic-wave solver then propagates a spark-discharge pulse from a candidate emitter position in the corridor and records $\sigma_1$ everywhere.

Selected high-$S$ cases are passed to a quasi-static fracture finite-element analysis. This checks whether the calcium stress pattern predicts a through-crack across the fragment, which is the spinal analogue of the hinge effect, rather than surface fissuring.

Geometry sources are as follows. The vertebral anatomy of the 63 volunteers in A large, paired dataset of robotic and handheld lumbar spine ultrasound with ground-truth CT benchmarking provides normal L1–L5 shape variation. Because those volunteers are healthy, calcified fragments and osteophytes have to be synthetically inserted, with parameters for size, calcified arc, thickness, and distance to dura. That makes this a development set, not a clinical one. The MDCT protocol supplies density-to-stiffness mapping. A genuinely external set must be assembled at a second center: retrospective CT of patients with calcified lumbar disc herniation, ideally on a different scanner vendor. Nothing in the file provides it, and it is the most important data-collection task.

The vascular source domain comes from OCT-derived ring geometries in the style of the Disrupt CAD-III models. As a test of the geological hypothesis, a second source domain uses layered impedance models derived from the Parihaka 2D velocity models, rescaled to biological impedance ranges.

Second, we train a Fourier neural operator, following the velocity-continuation design. It takes a tissue-property map plus an emitter position and outputs the peak-stress field. Training sets differ by experiment, as described below.

Third, the surrogate is used for the one task where direct simulation is too slow: propagating uncertainty. CT-to-stiffness mappings and calcium properties are uncertain. For each patient we sample property fields from that uncertainty and sweep emitter positions. The output is a per-patient map of the probability that $S$ exceeds a safety margin while the calcium stress exceeds a fracture threshold.

The dural damage threshold is not established in the file. We therefore treat it as a swept parameter and report conclusions across a range rather than at one invented value.

Baselines

The naive reference point is the "transfer by analogy" assumption: lithotripsy is selective in arteries, so it is selective in the spine. Concretely, this is a one-dimensional layered transmission model. It computes reflection and transmission along the straight ray from emitter through fragment to dura, using the same tissue properties as the full solver, at negligible cost. It is beaten if its predicted $S$ misclassifies safe versus unsafe configurations, relative to the full solver, at a rate that would change a surgical decision. We will report this as disagreement rate and AUC of the one-dimensional $S$ against full-solver labels. If the one-dimensional model agrees closely, the project's main claim is false and we say so.

The strongest existing method for producing stress fields is the direct numerical solver itself. Every surrogate is scored against it on identical geometries and property samples. The comparison measures wall-clock cost to reach a fixed Monte Carlo standard error on the per-patient unsafe-probability estimate. The surrogate beats the solver only if it reaches that precision at lower total cost, with training-data generation counted.

The strongest learned competitor is an in-domain surrogate. This is an operator trained only on spine geometries from the same center, exactly as the Parihaka operator was trained on its own survey. It receives the same architecture, the same simulation budget, and the same hyperparameter search. On the clinical question, the reported number to beat is absent: we found no reported result on Lacuna for neural-operator field accuracy or for any computational prediction of spinal lithotripsy selectivity. The velocity-continuation paper reports millisecond inference and qualitative phase/amplitude agreement, but no error metric in the file. The closest quantitative transport benchmark is the random forest's external $R = 0.71$ (from $0.95$ in training) for Young's modulus. We use it only as a calibration of how much degradation to expect, not as a target on our task.

The ablation that isolates the key ingredient is cross-medium pretraining. We train three operators: on spine only; on vascular plus spine; and on vascular plus geological layered media plus spine. The spine simulation budget is held fixed, and all three are evaluated frozen on the external center.

Why not just…?

Why not just run the full solver on every patient and skip the surrogate? For deterministic feasibility on a few dozen patients, this is the right tool, and the first deliverable uses it. The surrogate earns its place only for uncertainty quantification over material properties and emitter sweeps, where thousands of runs per patient are needed. Experiment three settles this by counting solver runs. If uncertainty over CT-derived stiffness barely moves $S$, few samples suffice, the solver wins, and the surrogate is dropped from the clinical pipeline.

Why not just compute impedance mismatches analytically? This is the most serious threat. Selectivity is, after all, an impedance argument. Our hypothesis is that the bony canal refocuses reflected waves off-axis, which a one-dimensional ray cannot see, and that this geometric effect is what makes the spine differ from the artery. The first experiment tests that hypothesis directly. If the one-dimensional model's $S$ matches the full solver, the simpler method wins.

Why not just retrain a survey-specific surrogate per center, as geophysicists do? If retraining needs only a few hundred simulations, the cost is modest, and cross-domain transport is an academic concern. The learning-curve experiment measures how many target-center simulations the in-domain operator needs to match a frozen cross-domain operator. If the answer is small, we recommend per-center retraining and say that cross-medium pretraining is unnecessary.

Why not just go to cadaver bench tests with an existing coronary lithotripsy catheter? Bench work is the eventual arbiter. However, a cadaver sample covers few anatomies, cannot measure dural stress fields, and cannot separate geometry effects from tissue-property effects. Simulation tells us which anatomies and emitter positions to test, and what the cadaver result must look like to confirm or refute the model.

Proposed Experiments

The first decisive check is the full solver against the one-dimensional model. It covers synthetic-fragment spine geometries spanning fragment size, calcified arc (90° and 270°, mirroring the vascular study), distance to dura, and bone density. We report the distribution of $S$, the fraction of configurations below a swept safety margin, and the agreement between the one-dimensional and full models. In the vascular control geometries, the full solver should reproduce favourable selectivity. If it does not, the simulation pipeline is wrong.

The second check is transport. Operators are trained on development-center spine geometries, with or without vascular and geological pretraining, then frozen and evaluated on the second-center calcified-disc CT cohort. Metrics are relative field error in $\Omega_{\text{calc}}$ and $\Omega_{\text{dura}}$, error in $\log S$, and AUC plus calibration for classifying unsafe configurations. The gap between internal and external performance is reported explicitly. We also report a learning curve showing how many target-center simulations close that gap.

The third check is uncertainty quantification. For each external patient we compare the surrogate's unsafe-probability estimate against a smaller full-solver Monte Carlo reference, and record the cost to reach matched precision.

The fourth check is mechanism. For the highest-$S$ configurations we run fracture finite-element analysis to test whether predicted stress yields through-fractures of the fragment. We also test whether fracture length or position, by analogy with the vascular findings, governs fragment release.

The project succeeds in either of two ways. It may produce a validated, transportable map of when spinal lithotripsy would be selective. Or it may produce a clear negative result, showing that canal geometry destroys the selectivity that arteries enjoy.