Some spine operations reach the lumbar discs from the side of the abdomen, behind the abdominal lining, instead of from the back. Oblique lumbar interbody fusion (OLIF) is one of these. It spares the back muscles and joints, but the instrument path runs within millimetres of the aorta and the inferior vena cava, the body's largest vein. Navigation software and the newer learned trajectory planners judge a plan by how far the planned path stays from critical structures. Surgeons do not execute plans exactly. In a CT-navigated cohort of 140 screws, the placed screw tips ended up 5.5 ± 2.7 mm from the plan and the axes 6.3° ± 3.6° off, while the navigation system's own registration error was only 0.87 mm. All of those screws were still graded clinically acceptable (CT-Navigated Spinal Instrumentations). Inside bone, that drift is absorbed. In a corridor bounded by a compliant, displaceable great vessel, it may not be.
We propose a procedural benchmark generator for oblique and lateral lumbar access planning. It synthesizes many thousands of anatomically parameterized corridors with exact ground truth. These include rare vascular layouts such as a left-sided vena cava. Each candidate plan is scored by its probability of vascular contact and excessive retraction strain once realistic execution drift is applied, not by its clearance on paper. The claim is a measurement claim. Ranking planners by plan-time clearance on a small fixed set of scans gives a different and less stable answer than ranking them by drift-aware vascular risk over a large procedurally generated set. The drift-aware ranking is the one that predicts physical outcomes on a phantom.
Core Claim
The three directions play different roles here. Spinal preservation supplies the application: OLIF and related lateral corridors, which a meta-analysis credits with shorter operations, less blood loss and better disc-height restoration than posterior minimally invasive fusion (Meta-analysis of MIS-TLIF versus OLIF). Iatrogenic vascular injury supplies the failure model. Vessels fail when the path drifts into them, and also when ordinary static retraction exceeds a lowered tolerance in friable tissue, as in retractor-associated rupture in neurofibromatosis (Spontaneous aortic rupture). Dynamic benchmark generation supplies the method. Verifiable instances are generated from parameters, rare regions of the parameter space are held out, and enough instances are drawn to separate methods statistically. NetArena showed that agent confidence intervals overlapping by 85% at 100–150 queries stopped overlapping at thousands (NetArena).
The specific hypothesis has three parts. First, a planner's vascular risk under drift is not a monotone function of its planned clearance, because drift is anisotropic and level-dependent (larger at S1 in the navigation cohort) and because retraction moves the vessel. Second, rankings computed on fewer than a few hundred cases are unstable. Third, planners tuned on typical anatomy fail disproportionately on held-out vascular variants. If drift-aware scoring and plan-clearance scoring agree on rankings, or if a small real cohort ranks planners just as reliably, the claim fails.
Research Program
The generator builds a lumbar segment (L2–L5 discs) from parameters: disc height and collapse, osteophytes, coronal tilt, psoas position and bulk, and aortic and caval position, diameter and bifurcation level. A discrete anomaly switch covers configurations such as the transposed vena cava that forced a change of patient side and approach in A patient with left-sided inferior vena cava who received OLIF. Because we own the geometry, signed distances to every vessel and disc-coverage targets are computed exactly. This plays the role that SAT solvers play in BeyondBench, which uses them to guarantee ground truth for its procedurally generated reasoning tasks (BeyondBench). Our domain is continuous geometry rather than symbolic puzzles.
Each instance can optionally carry a CT-like rendering. Planners that consume images then face realistic appearance. A text-guided 3D CT synthesizer such as MedSyn is one candidate texture layer. Ground truth always comes from the parametric geometry, not from the synthesizer.
A plan $\pi$ consists of a patient side, an entry point, a trajectory to the disc, and a retractor blade placement. We score it by drift-aware risk:
$$R(\pi) = \Pr_{\epsilon \sim D_\ell}\big[\, d(\pi \oplus \epsilon,\, V(\pi)) < 0 \ \text{ or } \ s(\pi) > \tau_\phi \,\big].$$
The symbols are defined as follows. $\epsilon$ is an execution error, combining tip offset and axis rotation, drawn from a level-dependent distribution $D_\ell$. Its moments are matched to the navigation cohort. $\pi \oplus \epsilon$ is the executed path. $V(\pi)$ is the vessel geometry after displacement by the retractor. $d$ is signed clearance. $s(\pi)$ is retractor-induced vessel strain from a simple linear-elastic surrogate. $\tau_\phi$ is a strain tolerance that shrinks with a friability parameter $\phi$.
Alongside $R$ we report task success, meaning cage footprint coverage of the disc. NetArena found that correctness and safety trade off (NetArena), and a planner that hugs the psoas and barely reaches the disc should not look good.
We will be explicit about two things we do not model. Delayed erosion after the adventitia (the outer protective layer of the artery wall) is stripped (post-hepatectomy complications) is a weeks-long biochemical process. We report only a proxy for it: the length of the instrument path in tangential contact with the vessel wall. Pulsation and bleeding are also out of scope. We will not claim that the benchmark measures either.
Splits are defined over generator parameters, not over instances. Learned planners train on typical-anatomy regions. Testing covers held-out regions: the left-sided vena cava, low bifurcation, high friability, and collapsed discs with large osteophytes. This guards against the generator-specific overfitting that a fixed seed list would allow.
Baselines
The naive reference is a fixed textbook plan: right lateral decubitus position, a standard entry window between the psoas and the vessels, and a trajectory perpendicular to the disc. It is applied identically to every instance. Any planner worth deploying must beat it on drift-aware risk without losing disc coverage. On the left-sided vena cava split we expect it to fail often. That is exactly the situation the case report resolved by hand.
The strongest existing approach to vascular avoidance is differentiable planning with a hinge-loss margin against segmented vessels. Percept-Aware Surgical Planning for Visual Cortical Prostheses with Vascular Avoidance optimized electrode coordinates against a cerebrovascular atlas this way and reported eliminating 100% of safety margin violations at a 300 µm margin, with a 1.7–4.4% SSIM cost. It is the strongest competitor because it already does what a skeptic would propose: optimize clearance directly. Its static, non-deforming assumption is the thing our metric tests. We port it to the lumbar geometry with millimetre margins swept over a grid. It receives the same instances and the same optimization budget as every other planner. It is scored on $R$ and on disc coverage.
The strongest learned approach is constrained reinforcement learning with a distance-based safety filter, the design of SafeRPlan. That system reported 98–100% safety on synthetic spines and 99% acceptable Gertzbein-Robbins grades on 35 volunteer vertebrae, but only for bony pedicle containment. We found no reported result on Lacuna for vascular-corridor planning. We reimplement its filter with vessels added as constrained boundaries. It trains on the training split with the same number of environment steps as our drift-aware variant and the same hyperparameter-search budget.
Two planners isolate the key ingredient, which is knowledge of drift and vessel displacement. The first is a margin-inflated planner: the hinge-loss planner with its margin set to a high quantile of $D_\ell$. The second is the same constrained reinforcement learning agent trained with $R$ as its constraint rather than plan-time clearance. Beating a baseline means lower $R$ at equal or higher disc coverage, with non-overlapping bootstrap intervals on the held-out splits. For the benchmark itself, the competitor is evaluation on a small fixed set of scans with plan-time clearance as the metric. Beating it means better agreement with phantom outcomes, described below.
Why not just…?
Why not just evaluate on a few hundred real contrast CT scans? Real scans have real anatomy, and we will use a small institutional set, which we would have to obtain, as an anchor. Two problems remain. First, the dangerous configurations are rare: a left-sided vena cava appears as a case report, not as a cohort. Second, NetArena's finding that confidence intervals overlap 85% at 100–150 instances suggests a few hundred scans may not separate planners whose risk differs by a few percent. We test this directly. We plot ranking stability against sample size on both the procedural set and the real set. If the real set ranks planners stably and matches phantom outcomes as well as the procedural set does, the generator adds little, and we will report that.
Why not just inflate the safety margin by the measured drift? This is the real threat. It may perform as well as anything on typical anatomy. It is included as a baseline precisely so the benchmark can show whether it suffices. Our expectation, which is unverified, is that a uniform margin either wastes the corridor or fails. Drift is anisotropic and level-dependent, and retraction moves the vessel toward the path, so a uniform margin either closes narrow corridors entirely (losing disc coverage) or under-protects on the drift axis. If the margin-inflated planner matches the drift-aware reinforcement learning agent on held-out splits, the conclusion is that a simple rule is enough. That would itself be a useful, publishable result from the benchmark.
Why not just extend SafeRPlan's distance filter to vessels? That is one of our baselines. A filter that guards plan-time distance inherits the gap between the plan and what is executed. The navigation cohort shows that gap is real even when registration is sub-millimetre. The comparison between the vessel-filtered agent and the same agent constrained on $R$ answers this directly.
Why not just simulate with a full finite-element tissue model? It would be more faithful but far too slow to generate thousands of instances. We instead calibrate the linear surrogate against finite-element runs on a subset of instances and report the rank correlation. Low correlation would invalidate the strain term.
Proposed Experiments
The first experiment tests whether the metric matters. We compute Kendall rank correlation between planner rankings under plan-time clearance and under $R$, separately for each split. The claim requires a substantial disagreement concentrated on the anomaly and friability splits.
The second tests whether scale matters. We bootstrap the confidence intervals of the planner differences at N = 100, 300, 1,000 and 10,000 instances, and measure overlap in the way NetArena did.
The third is the decisive physical check. We 3D-print lumbar segments from about twenty generated instances spanning the splits, including two left-sided vena cava layouts. Compliant tubes stand in for the vessels, as the latex vessels of the dVRK shunt work did (Automating Vascular Shunt Insertion with the dVRK Surgical Robot). Each planner's top plan is executed with a tracked instrument by operators of varying experience. We record realized contacts and tube displacement. The benchmark earns its keep only if $R$ predicts phantom contact rates better than plan-time clearance does. We will also report the realized drift so that $D_\ell$ can be checked against what the phantom produces rather than assumed.
References
- CT-Navigated Spinal Instrumentations–Three-Dimensional Evaluation of Screw Placement Accuracy in Relation to a Screw Trajectory Plan
- Meta-analysis of minimally invasive transforaminal lumbar interbody fusion versus oblique lumbar interbody fusion for treating lumbar degenerative diseases
- Spontaneous aortic rupture in a patient with neurofibromatosis type 1
- NetArena: Dynamic Benchmarks for AI Agents in Network Automation
- A patient with left-sided inferior vena cava who received oblique lumbar interbody fusion surgery: a case report
- BeyondBench: Contamination-Resistant Evaluation of Reasoning in Language Models
- Hepatic Pseudoaneurysm Ruptured into the Jejunal Limb after Hepatectomy for the Treatment of Gall Bladder Carcinoma
- Percept-Aware Surgical Planning for Visual Cortical Prostheses with Vascular Avoidance
- SafeRPlan: Safe deep reinforcement learning for intraoperative planning of pedicle screw placement
- Automating Vascular Shunt Insertion with the dVRK Surgical Robot