Spine surgeons who aim to preserve the spine's own structure make many small removal decisions. In one decompression, a surgeon may take the ligamentum flavum (the thick ligament behind the spinal canal) at one level, shave a quarter of a facet joint (the small paired joints at the back of each vertebra) at the next, and leave the supraspinous ligament intact throughout. Each decision may be safe alone while the combination destabilizes the spine. Planning software ignores this. Trajectory tools treat vertebrae as rigid bodies and tissue as either free space or obstacle. Clinical evidence only says that particular whole procedures preserve stability. This note proposes a planner that represents surgical plans as a partially ordered set, where one plan sits below another if it removes less tissue. It uses one deliberately weak rule: the planner is told explicitly which combinations of individually safe plans are themselves known to be safe, and it never assumes the rest.
The mathematical object is a marked structure. This is an ordered set in which "least upper bounds" (here, the union of two plans) are only required to behave well over a designated family of admissible subsets, not over all of them. A completion construction then embeds it into a structure where every union exists, without changing the meaning of the original admissible plans. The practical payoff is not elegance. A planner gets two sound inference rules it can use to avoid running expensive biomechanical simulations. Anything below a stable plan is stable, and anything above an unstable plan is unstable. It also gets a precise account of where those rules fail: at the union of two stable plans, which may or may not be stable. We test whether this account cuts the simulation budget needed to map the safe region of multi-level resection plans, and whether the extra plans the completion adds cause real planning errors.
Core Claim
Reflectors to quantales defines a marked quantale as an ordered semigroup paired with a family $\mathcal{A}$ of admissible subsets. Joins must exist over those subsets, and multiplication must distribute over them. The paper proves that the "$\mathcal{A}$-ideals" construction $\mathrm{Id}_{\mathcal{A}}(S)$, which collects lower sets closed under admissible joins, is the free completion and an order embedding. The same paper flags an open problem: this completion need not preserve joins that exist in $S$ but are not marked.
Our claim is that this open problem is exactly the failure mode of compositional surgical planning. The construction works as follows.
The elements of $S$ are concrete resection plans along a lumbar segment range, ordered by inclusion. Each level contributes a product of short chains, such as facet resection fractions 0, 25, 50 or 75 percent per side, flavum removed or kept, and interspinous or supraspinous ligament removed or kept. The marked family $\mathcal{A}$ contains the sets of plans whose union has been certified stable by simulation. The multiplication composes plans on adjacent spinal segments. Its distributivity over admissible joins is a locality hypothesis: merging plan envelopes at one level commutes with attaching a plan at the neighbouring level.
This gives an empirically falsifiable law. The planner searches in the completion, where unions always exist and search is easy. Any returned element that is not the image of a real certified plan is flagged as a "spurious join". This is the computational face of the unpreserved-joins problem. The claim is that a planner built this way certifies a larger share of the truly safe plan space per simulation than unstructured or purely monotone alternatives, and never certifies an unsafe plan.
The motivating clinical evidence is that selective sparing, not total avoidance, is what preserves stability. Semi-Circumferential Decompression removes the entire ligamentum flavum en bloc while sparing facets and the supraspinous ligament. In 19 Grade I spondylolisthesis patients, slip went from 10% to 12.2% with no significant change. Corridor approaches like those in the OLIF vs. MIS-TLIF meta-analysis gain stability by not disturbing posterior elements at all. Neither paper says which mixtures of partial resections across levels remain safe. That is the question here.
Research Program
The anatomical substrate comes from VerSe, which provides CT scans with voxel-level vertebra labels. The MRI route uses SPINEPS segmentation followed by the rule-based landmarking of Graf et al., implemented in TPTBox. That method extracts ligament and muscle attachment points and vertebral orientations with 1.72° mean angular error. It was built to feed multibody biomechanical simulations, which is how we will use it.
For each subject's lumbar spine, we build a multibody model with ligaments as attachment-point springs and facets as contact surfaces. A plan is applied by deleting or weakening the corresponding elements. Instability is scored as the increase in simulated segmental translation under a flexion/extension load. The pass threshold is anchored to the clinically tolerated change in the SCD cohort.
The planner maintains three things: the certified-stable set, the certified-unstable set, and the marked family $\mathcal{A}$. It infers stability downward from certified plans and instability upward. A union is only certified once it is simulated or implied by a certified plan above it. It chooses the next simulation query by maximizing how many uncertified plans would be resolved either way.
Because the completion turns every union into a well-defined object, the planner can rank candidate plans by decompression value inside the completion. Decompression value is measured as canal and foraminal area gained, using the cross-sectional measures that Abbati et al. found most predictive of symptomatic levels. The planner then checks whether the winner is a principal ideal, meaning a real certified plan. If it is not, it is resolved by a simulation query rather than accepted.
We deliberately do not use the twisted-identity (BiHom) algebras or the equational-systems machinery from the file. The BiHom constructions in Asif and Wang have no semantics we can tie to measured spinal coupling. For label shifts, the equational approach of Fiore and Hur would reduce to an ordering constraint that ordinary sequence decoding already enforces. Neither would survive the simpler alternatives a reviewer would propose.
Baselines
The naive reference point is per-level independent thresholds, the implicit rule in current practice. Each level's plan is accepted if it is stable when simulated alone, and multi-level plans are accepted if every level passes. This uses one simulation per level-plan and so has the smallest budget. It is beaten if it certifies any plan the full simulator calls unstable, which we expect on adjacent bilateral facet resections. The planner must certify at least as many safe plans per simulation without that error.
The strongest practical competitor is a monotone classifier, such as gradient-boosted trees with monotonicity constraints on every resection variable. It is trained on the same simulated plans, chosen by the same uncertainty-driven query rule and the same simulation budget, and its hyperparameters are tuned on held-out subjects with an equal number of tuning runs. It captures the downward-closure half of our structure but generalizes freely across unions. Beating it means a lower false-safe rate at equal recall of truly safe plans, or equal error with fewer simulations. No reported result was found on Lacuna for stability prediction from multi-level resection plans, so this comparison is new rather than benchmarked.
A constrained integer program receives the same queried simulation outcomes. Stability is encoded as linear constraints learned from those outcomes: hard per-element limits plus pairwise interaction terms. This is the "just optimize" baseline. It is run with the same data and budget.
The key ablation is our planner with the marked family replaced by "all unions of stable plans are stable". That is a full quantale rather than a marked one. It isolates whether refusing to infer unions is what prevents unsafe certifications. A second ablation drops the spurious-join check, which measures the cost of the unpreserved-joins problem directly.
Anatomical accuracy feeds every method equally. For context, the input stage has reported performance of a 96.6% identification rate and 91.72% Dice for the VerSe 2020 winning model (VerSe). Trajectory transfer under deformation has a 0.94 mm mean RMSE reported for functional maps (Pedicle Drilling Planning Transfer). Neither number measures stability, and we do not claim to compete with either.
Why not just…?
Why not just use a constrained optimizer with pairwise interaction terms? This is a real threat if instability is truly pairwise and additive. The integer program is in the baseline set for that reason. The order-theoretic planner should win only when unions of three or more partial resections are unstable even though every pair is stable. That is a higher-order interaction a pairwise model cannot express without enumerating it. We will measure how often such triples occur in simulation. If they are rare, the honest conclusion is that the integer program suffices, and the algebra serves only as documentation of its soundness.
Why not just train a monotone classifier and skip the algebra? Monotonicity is most of the inductive bias, and we expect this baseline to come close. The difference is that a classifier smoothly extrapolates to unions it has never seen, while the marked structure refuses to. In a safety setting, one false-safe certification matters more than many conservative refusals. The decisive comparison is therefore the false-safe rate at fixed recall. If the classifier matches us there, the marked structure buys nothing.
Why not just use deep networks on images to predict outcomes directly? Abbati et al. found that CNNs and autoencoders on raw MRI reached 69.8% and 70.6% AUC for symptomatic-level prediction, against 85.4% for a random forest on 22 structured radiological features. Their data volume is similar to what any surgical stability study could assemble. This supports building on structured anatomical quantities and simulation rather than learning stability from pixels.
Why not convex safe-corridor planning from robotics? Safe-corridor methods such as relative safe flight corridors decompose free space into convex regions with binary occupancy. Our question is not geometric reachability but which removals of load-bearing tissue are tolerable. That is a set-valued and ordered decision that has no occupancy representation.
Proposed Experiments
The primary experiment uses 20 to 40 lumbar spines from VerSe, plus MRI subjects processed through SPINEPS and TPTBox. Each spine gets a resection lattice over L3–S1 with roughly a few thousand plans, small enough that every plan can be simulated once to produce ground truth. All methods then run under budgets of 5, 10, 20 and 40 percent of the lattice. We report the false-safe rate, the recall of truly safe plans, and the decompression value of the best certified plan.
The first decisive check is the locality law. For random pairs of certified plan sets at one level combined with a fixed plan at the adjacent level, we test whether simulated stability of "combine then merge" equals "merge then combine". Systematic failure falsifies the distributivity hypothesis and tells us the coupling distance along the spine.
The second check counts spurious joins. These are completion elements that rank highest but correspond to no certified plan. We report how often accepting them, as the full-quantale ablation does, yields unstable plans.
The third is a clinical sanity check. An SCD-type plan (full flavectomy with facets and supraspinous ligament spared) should be certified safe. A laminectomy with bilateral 50% facetectomy over two levels should not. A simulator that disagrees with the SCD cohort is recalibrated before any comparison is trusted.
References
- Reflectors to quantales
- Semi-Circumferential Decompression: Microsurgical Totalen-blocLigamentum Flavectomy to Treat Lumbar Spinal Stenosis with Grade I Degenerative Spondylolisthesis
- Meta-analysis of minimally invasive transforaminal lumbar interbody fusion versus oblique lumbar interbody fusion for treating lumbar degenerative diseases
- VerSe: A Vertebrae labelling and segmentation benchmark for multi-detector CT images
- Rule-based Key-Point Extraction for MR-Guided Biomechanical Digital Twins of the Spine
- MRI-Based Surgical Planning for Lumbar Spinal Stenosis
- Rota-Baxter operators and Loday-type algebras on the BiHom-associative conformal algebras
- On the construction of free algebras for equational systems
- Pedicle Drilling Planning Transfer for Spine Surgery Using Functional Map Correspondences
- Fast Trajectory Planning for Multiple Quadrotors using Relative Safe Flight Corridor