Planning Calcified-Artery Cracking on a Crack-Topology Graph

When an artery is narrowed by a hard ring of calcium, inflating a balloon inside it often fails. The ring behaves like a rigid tube, and the balloon either cannot open it or opens it only at pressures that tear the vessel. Clinicians therefore crack the calcium first. Intravascular lithotripsy (IVL) does this by firing acoustic shockwaves from a balloon catheter; the shockwaves pass through soft tissue and fracture calcium. A cheaper trick, the parallel wire technique, does it by laying a stiff guide wire beside the balloon so the inflation force concentrates on one line, much like a wedge. Finite-element simulations of these procedures point to a striking regularity: what predicts how much the lumen opens is mostly whether the calcium ring has been broken through, and much less the exact shape of the cracks. Sogbadji et al. found that a crack running fully through the plaque turns the ring into a hinge and gives up to 51% relative lumen gain in large lesions, against 16–25% for non-penetrating cracks of the same length. They also found that crack length matters, while crack width and crack position barely do.

We propose to make that regularity the representation. Each calcified cross-section, and each short segment of vessel, is encoded as a small graph: calcium arcs are nodes, and cracks are edges labelled by depth, with a through-crack cutting the ring. A neural network that reads this graph is trained on finite-element simulations to predict lumen gain and peak strain on the healthy wall. We then use this fast predictor to plan sequences that mix tools, such as shockwaves first and then a wedge wire placed at a chosen angle. The tools have very different effects, so the planner must decide which one carries how much of the risk to the wall. The claim is specific. For lesions with crack configurations never seen in training, the graph predictor should rank candidate procedures correctly where image-based and flat-feature predictors fail, and plans chosen this way should reach a hinge with lower wall strain than single-tool clinical defaults.

Core Claim and Where Each Direction Sits

Focal stress concentration and lithotripsy supply the physics and the application. Lithotripsy produces diffuse, often non-penetrating fissures. The parallel wire concentrates force onto a single line and has worked clinically: Horita et al. report two fistula stenoses that resisted 12 atm balloon angioplasty but opened at 10–12 atm with the wire. Yamaya et al. report 100% 24-month patency in 9 wire-assisted lesions against 41.2% with conventional dilation, though from a small, non-randomized cohort. Nobody has asked whether the two tools are complementary. Diffuse shockwave fissuring could create the weak point that lets a wedge produce a through-crack at low pressure.

The "structural representation gap" is the methodological lens. Work on formula encoders shows that flat sequences lose hierarchy and break under trivial surface changes, while graph and tree encoders keep it. FormulaNet embeds formulas as graphs for premise selection; FORTE embeds them as trees; TeLoGraF feeds a permutation-invariant graph encoding of temporal-logic specifications to a trajectory generator. We reuse the TeLoGraF-style idea of an explicit structural encoder whose output does not change when equivalent parts are relabelled. Here a rotated lesion, or reordered but equivalent arcs, should give the same prediction. These papers motivate the hypothesis, but none touches physical fracture.

Heterogeneous agent coordination enters as a constraint on planning rather than as a learning algorithm. Wang, Ames and Egerstedt split collision-avoidance responsibility among robots in proportion to their agility instead of 50/50. We borrow that principle to split a wall-strain safety budget between tools in proportion to how selective they are. DrIGM proves that training each agent against its own worst case breaks team coordination, whereas a single shared worst case does not. We test the analogous prediction in our setting: giving each tool its own worst-case tissue margin should compound into under-treatment, while one shared worst case over patient tissue parameters should not. We deliberately do not use multi-agent reinforcement learning, for reasons given below.

Research Program

The first stage builds the data. We extend the 22-model Abaqus/Explicit suite of Sogbadji et al., along with its companion in silico assessment, to roughly 2,000 two-dimensional cross-sections and about 200 short three-dimensional segments. The extension varies calcium arc from 60° to 360°, uses one to three separate calcium arcs (eccentric lesions), varies thickness along the arc, and includes zero to four cracks of varying depth and length. Each model is expanded to 14 atm, matching their protocol. A model is accepted only if the extended suite first reproduces their two reported trends: through-crack versus non-penetrating lumen gain (51% versus 16–25%), and their global benefit metric of 192 for 270° lesions versus 144 for 90° lesions.

The second stage builds the predictor, which we call a surrogate because it stands in for the slow finite-element solver. Nodes carry arc angle, mean and minimum thickness, and axial extent. Edges carry crack length and depth, plus a binary through-cut flag that changes the ring's connectivity. A message-passing network, the kind that updates each node from its neighbours, predicts relative lumen gain and peak healthy-wall strain.

The third stage builds tool-effect models and a planner. The wire acts on the graph as a directed operator: at a chosen angle, it adds a crack whose probability of going through depends on local thickness and on existing fissures. Lithotripsy acts as a stochastic operator: it adds shallow fissures whose density falls with distance from the emitter. Its parameters are fitted to crack statistics where OCT data such as Disrupt CAD-III pullbacks are accessible, and are otherwise set as explicit, sensitivity-tested assumptions. This is the weakest link, because no paper in our sources models shockwave-driven fracture dynamically. Sogbadji et al. pre-cut their cracks too. We therefore report every planning result under a range of lithotripsy parameters, not one setting.

The candidate set of procedures is small: balloon alone; wire alone at k angles; lithotripsy at a few pulse counts; lithotripsy then wire; and wire then lithotripsy. The order is a real variable, since the sequential world models work SeqWM shows that outcomes of coordinated agents depend on who acts first. The planner enumerates this set and scores each plan by expected lumen gain. It requires peak wall strain to stay under a threshold for the worst case over a sampled set of tissue parameters, shared across both tools.

Baselines

The naive reference is balloon-only dilation, simulated at 12 and 14 atm, which corresponds to the conventional arm in Yamaya et al.. Any plan must beat it on lumen gain at equal or lower wall strain. The single-tool clinical defaults, lithotripsy alone and wire alone, are the practical comparators for the planning stage.

The strongest existing method is direct finite-element simulation as done by Sogbadji et al.. It serves as the oracle that every surrogate is scored against, and its plan choices define zero regret. Lacuna holds no reported result for a learned surrogate of post-crack lumen gain or for automated tool-sequence planning, so the comparison numbers will be generated here. The physical anchors we must reproduce are the reported 51% versus 16–25% relative lumen gain and 192 versus 144 global benefit.

Among learned competitors, the strongest is a gradient-boosted regressor on hand-built features: maximum arc, minimum thickness, number of through-cracks, and total crack length. It encodes Sogbadji's own findings directly. We also run a multilayer network on the same flat parameter vector and a convolutional network on rasterized cross-section images.

All learned models get identical training splits, the same 50-trial hyperparameter search, and comparable parameter counts. The key ablation keeps the graph network but removes the through-cut edge semantics, so cracks become plain node features and the ring's connectivity never changes. It isolates whether the topology itself is doing the work.

We fix the target for "beating" in advance. On the held-out topology splits, the graph surrogate's error on lumen gain should be at most half that of the best flat baseline, and its plan-ranking correlation with the oracle should be clearly higher. If the gap is smaller than the spread across random seeds, the representation claim fails.

Why not just…?

Why not just regress on a few hand-built features? This is the real threat, because the lumen-gain results already reduce lumen gain to roughly "is there a through-crack, and how long are the cracks." On single-crack, single-arc lesions we expect such features to match the graph model, and we will say so. The question is whether they hold up on two or three cracks across several separate arcs. There, the count of load-bearing fragments and which arcs are spanned are not summarized by the features. The held-out-topology split settles it. If the features keep pace there, we drop the graph and plan with the regressor.

Why not just run finite elements for every candidate plan? For one lesion with ten deterministic plans, a few hours of offline simulation may be acceptable. A surrogate becomes necessary under uncertainty: the shared worst-case check needs hundreds of tissue-parameter samples per plan, and lithotripsy outcomes are stochastic. We report the wall-clock cost of oracle planning against surrogate planning at matched decision quality.

Why not just use multi-agent reinforcement learning to coordinate the tools? With two tools, a handful of angles, and two possible orders, exhaustive search over a fast surrogate is exact and easy to audit. Reinforcement learning would add sample cost and opacity for no gain. We keep only the principles from coordination work: asymmetric safety responsibility from Wang, Ames and Egerstedt and shared versus per-agent worst cases from DrIGM.

Why not just inflate a non-compliant balloon to high pressure? The balloon-only baseline tests this directly. The clinical record already points against it: 12 atm failed in both of Horita et al.'s cases, and conventional dilation had 41.2% patency in Yamaya et al.. We measure the healthy-wall strain it needs to match a hinge's lumen gain.

Proposed Experiments and Decisive Checks

The first check is generalization. Train on lesions with at most one crack and arcs of 90° and 270°, then test on two to three cracks, multiple arcs, and 180° and 360° arcs. Metrics are mean absolute error on relative lumen gain and on peak wall strain, plus Spearman correlation between surrogate and oracle plan rankings. The cut-edge ablation and the hand-feature regressor are the two comparisons that matter.

The second check is planning. On 100 held-out lesions, every chosen plan is re-run in finite elements. We report the fraction of lesions that reach a through-crack hinge within the strain limit, and the regret against the oracle-best plan. The comparison is against balloon-only, lithotripsy-only, wire-only, and a fixed "lithotripsy then wire at the thinnest point" heuristic.

The third check tests complementarity. The hypothesis predicts that lithotripsy followed by the wire produces through-cracks at lower simulated pressure than either tool alone, most strongly in thick 270°-plus lesions. If it does not, under any plausible lithotripsy parameters, the mixed-tool premise is wrong and we will report that.

The fourth check tests robustness. Plans chosen under per-tool worst-case margins are compared with plans chosen under a shared worst case. The DrIGM-style prediction is that per-tool margins produce systematically sub-therapeutic plans, meaning lower lumen gain at no safety benefit.