When a surgeon removes a tumor, cutting around it is only part of the job. The tumor also has to come out in one piece. Grasping, pulling, or squeezing it can tear its outer boundary (the capsule) and release live cancer cells, which then implant elsewhere. Clinical evidence for this is strong. Breaking a tumor up during removal speeds recurrence (Borella et al.). Removing bladder tumors intact instead of chipping them away lowered 1-year recurrence from 38.1% to 28.5% (En bloc resection of bladder tumour). Handling a mass before its blood supply is controlled pushes tumor cells into the circulation (Pan et al.). Current autonomous surgical planners ignore this failure mode. They keep the tool a fixed distance from the tumor, but "tool is far from the tumor" does not imply "tumor is not being torn": a retraction several millimetres away can still stretch the capsule past its breaking point.
We propose to make capsule strain, the fractional stretch of the tumor's boundary, the quantity the robot's safety layer constrains. A soft-tissue simulator predicts how the capsule will stretch under the next tool motion. A fast filter then edits the planned motion so that predicted strain stays below a rupture threshold, with a safety cushion that grows with the simulator's uncertainty. The central difficulty is the sim-to-real gap. Predicted strain depends on tissue stiffness, which varies across organs and pathologies, and existing real-to-sim tissue tracking assumes that stiffness is known. So the key ingredient is estimating stiffness online from the tissue's observed motion and feeding that estimate's uncertainty directly into the constraint. We test this on the da Vinci Research Kit (dVRK), an open research version of the da Vinci surgical robot. The tests use tissue phantoms and ex-vivo organs containing capsules loaded with fluorescent beads, so that "seeding" becomes a countable outcome rather than a surrogate.
Core Claim
A safety filter that bounds predicted capsule strain, calibrated online to the specific tissue, will produce fewer capsule breaches and less bead spill during robotic en bloc dissection than two alternatives: geometric keep-out zones and tool-force caps. It achieves this while removing no more healthy tissue. The gain depends on online stiffness identification. The same constraint evaluated with a fixed nominal stiffness, or trained with randomized stiffness, will either breach more often or become conservative enough to stall the dissection.
The formal statement is short. Let $x_t$ be the tool pose, $\theta$ the tissue material parameters, $\hat\epsilon(x_t;\theta)$ the simulator's predicted peak capsule strain, and $\epsilon_c$ the rupture strain measured for the capsule material. The filter enforces the barrier
$$h(x_t) = \epsilon_c - \hat\epsilon(x_t;\hat\theta_t) - \kappa\,\sigma_\epsilon(x_t) \ge 0,$$
where $\hat\theta_t$ is the current stiffness estimate, $\sigma_\epsilon$ is the spread of predicted strain across the posterior over $\theta$, and $\kappa$ sets how many standard deviations of cushion to keep. A quadratic program then minimally alters the nominal dissection command so that $h$ decays no faster than a chosen rate. This is the control-barrier-function construction used by SSP. What we change is the thing being protected: SSP protects rigid geometric volumes, and we protect a mechanical state of deformable tissue.
Where This Sits Relative to Prior Work
Three surgical-autonomy systems motivate the gap. Each constrains geometry or perception, not tissue mechanics.
From Monocular Vision to Autonomous Action reconstructs airway tumors from endoscopic video and plans an electrocautery cut with a fixed 1.0 mm clearance. It reports 0.41 ± 0.16 mm registration and 2.223 mm cut-surface error on static ex-vivo tissue, and takes 20–40 minutes of offline processing.
SSP filters learned surgical policies at 3–10 ms using control barrier functions over cylinders and bounding volumes. It supplies the filtering method we reuse.
SURESTEP plans tool trajectories to reduce tracking uncertainty and does not model tissue at all.
None of the three can express "do not tear the capsule."
Two lines of work supply the deformable-tissue machinery and also constrain its limits. Real-to-Sim Registration with Position-Based Dynamics keeps a position-based dynamics simulator aligned with stereo point clouds on the dVRK. Position-based dynamics is a fast simulator that moves mesh particles to satisfy distance and volume constraints. The registration keeps tracking under occlusion, but only with fixed material parameters and no cutting. SurRoL shows that contact physics fidelity drives real-robot success: NeedlePick rose from 58% to 86% on the real dVRK when crude attachment was replaced by friction contact. We read that result as a warning that a strain constraint built on the wrong stiffness will fail in the same way.
On closing the gap, RGAT learns an action correction that makes the simulator's next states match reality. CAD²RL and Sa et al. instead randomize the simulator. We take the identification route rather than either of these. Our constraint needs a calibrated uncertainty on one physical quantity, and neither an action correction nor a policy made robust by randomization provides that.
APEX distills slow trajectory optimizers into fast policies. We use its idea only for the nominal planner and leave its unaddressed sim-to-real question to our filter.
Metabolic prognostic markers play a deliberately narrow role here. Signatures such as SQLE-based sarcoma risk (Hong et al.) or HYAL2 in glioma (hazard ratio 2.527, Yan et al.) predict recurrence. They are measured on bulk tissue, usually after surgery, and nothing in this literature shows that they predict susceptibility to seeding. We therefore do not claim that metabolism should shape the robot's path. We use a preoperative biopsy-derived risk score only to choose an operating point on the tradeoff our system exposes: a tighter strain budget $\kappa$ at the cost of slower dissection and wider margins. Whether that choice improves outcomes is a clinical question beyond this proposal.
Research Program
The program has four parts.
The first is building a measurable seeding proxy. We will fabricate tumor analogues as stiff silicone or alginate inclusions wrapped in a thin hydrogel capsule loaded with fluorescent microbeads. We will measure the capsule's rupture strain $\epsilon_c$ on a tensile rig, and embed the analogues in two settings: gelatin phantoms of graded stiffness, and ex-vivo porcine liver and bladder wall. After each trial, fluorescence imaging of the surgical field and the irrigation fluid counts escaped beads, and the specimen is inspected for breach. This turns "iatrogenic seeding" into an endpoint that can be scored the same way on every trial.
The second is estimating capsule strain online. We extend the position-based dynamics registration of Real-to-Sim Registration so that each tissue region's compliance is a latent variable. We fit it by comparing observed surface displacement against simulated displacement under the dVRK's commanded tool motion, holding a small ensemble of parameter samples as the posterior. Capsule strain is read off the edge lengths of the inclusion's boundary mesh. Cutting is handled in the simplest defensible way: the planned dissection path runs through healthy tissue outside the capsule, so we delete mesh constraints along executed cut segments, and we do not model fracture in general. This scoping is a stated limitation, and it is the reason we restrict ourselves to en bloc dissection rather than tumor debulking.
The third is filtering at control rate. The simulator ensemble runs at a few tens of hertz and exports, for the current state, the predicted strain and its gradient with respect to tool motion. A quadratic program linearized around that state runs at the 3–10 ms rate SSP reports and edits each command. The nominal planner is a 3D trajectory generator that follows the planned dissection surface. It is a sampling-based optimizer distilled to a policy in the style of APEX, but any planner can sit upstream.
The fourth is characterizing the tradeoff. Sweeping $\kappa$ traces a curve of breach rate against healthy tissue removed and against procedure time. This curve is the object a clinician would use to trade handling risk against margin width. It is also where a biopsy-derived risk score would set the operating point.
Baselines
The naive reference is the fixed-margin planner of Acar et al.. It uses the same reconstruction, a 1.0 mm geometric clearance around the segmented inclusion, and no tissue model. Retraction is scripted to be the same as in our condition. It shows what happens when seeding is ignored entirely. Its reported figures are 0.41 mm registration and 2.223 mm cut-surface error, with no reported breach or seeding numbers. It beats itself only on speed.
The strongest existing method is SSP's control-barrier-function filter (SSP) with the tumor plus a margin encoded as a keep-out volume for both instruments, running on the same dVRK, nominal planner, and filter rate. SSP reports raising violation-free GauzeRetrieve success from 12% to 68%, with a 0.00% collision rate on NeedlePick in simulation. It is the strongest competitor because it already guarantees that the geometric constraint holds. Any advantage we show must therefore come from constraining the right quantity, not from having a filter at all. We tune its keep-out radius over the same grid size we use for $\kappa$. We claim to beat it only if, at matched healthy-tissue removal, breach rate or bead count is lower with a confidence interval excluding zero across tissue stiffness levels.
A force-cap baseline limits the estimated tool-tissue force from dVRK joint torques below a threshold tuned on the same budget. It is simple and is what many practitioners would try first.
The ablation that isolates our key ingredient comes in two forms, both using the same strain barrier. One uses a single fixed nominal stiffness, as the registration paper assumes. The other trains and evaluates with stiffness sampled from the full phantom range, in the domain-randomization style, and takes the worst case over that range. Both run on identical trials.
No reported result was found on Lacuna for capsule breach rate or bead or cell spill under any autonomous surgical controller, so all breach-related comparisons are new measurements.
Why not just…?
Why not just widen the geometric keep-out zone? Distance does not determine strain. Our dissection requires one instrument to hold and retract the tissue next to the tumor, and that pull loads the capsule from outside any keep-out zone. Pan et al. describe seeding caused by handling, not by incision. The SSP baseline swept over radius will show whether a large enough radius matches our breach rate. If it does, it should cost measurably more healthy tissue and dissection time. If it matches us at equal tissue cost, our claim fails and we will report that.
Why not just cap tool force? This is the most serious threat. Force and capsule strain are related, but the relationship depends on where the force is applied and on the surrounding tissue's stiffness. The same force stretches a capsule in soft parenchyma much more than in stiff tissue. A single cap tuned for soft phantoms should stall dissection in stiff ones, and a cap tuned for stiff phantoms should breach in soft ones. The graded-stiffness phantom sweep tests exactly this. If one force threshold performs well across all stiffness levels, the simulator is unnecessary and we will say so.
Why not just randomize stiffness instead of identifying it? Randomization works for quadrotors (Sa et al.) because a robust policy can succeed under any plausible dynamics. A strain bound that must hold for the softest plausible tissue forces the robot to retract gently everywhere, which we expect to show up as incomplete dissections or long procedure times on stiffer tissue. The randomized ablation measures this cost directly. If identification buys little over it, the added machinery is not justified.
Why not just widen the margin for metabolically high-risk tumors? That lookup table is trivial to implement, and we do not claim to outperform it. It changes margin width, while our system changes how the tumor is handled. The two act on different failure modes and could be combined. We avoid claiming that metabolic signatures predict seeding, because the file contains no such evidence.
Proposed Experiments
The primary experiment has 30 trials per condition per stiffness level, across three gelatin stiffness levels and two ex-vivo organs. The five conditions (fixed margin, SSP keep-out, force cap, strain barrier with fixed stiffness, strain barrier with randomized stiffness) are compared with our identified-stiffness strain barrier. Every condition uses the same capsule batch, randomized in trial order. Outcomes are capsule breach rate, escaped bead count, healthy tissue removed, margin achieved, completion rate, and time.
A calibration check comes before any control results. We compare predicted capsule strain against strain measured from capsule markers tracked in stereo video. It is decisive in the following way: if the identified model's strain error is not clearly below the fixed-stiffness model's, the downstream comparison is uninterpretable, and we stop to fix the estimator.
A mismatch stress test deliberately misreports the initial stiffness by a factor of two to four. It checks that the uncertainty cushion widens and breaches stay rare while the estimate converges. That is the property that distinguishes a calibrated constraint from a merely tuned one.
The decisive outcome is the curve of breach rate against healthy tissue removed. If our curve lies below SSP's and the force cap's across stiffness levels, strain is the right quantity to constrain. If either baseline's curve overlaps ours, the simpler method is sufficient.
References
- Clinical and Histopathological Predictors of Recurrence in Uterine Smooth Muscle Tumor of Uncertain Malignant Potential (STUMP): A Multicenter Retrospective Cohort Study of Tertiary Centers
- En bloc resection of bladder tumour: the rebirth of past through reminiscence
- Large Volume Osteosarcomas of the Femur Treated with Total Femoral Replacement
- SSP: Safety-guaranteed Surgical Policy via Joint Optimization of Behavioral and Spatial Constraints
- From Monocular Vision to Autonomous Action: Guiding Tumor Resection via 3D Reconstruction
- SURESTEP: An Uncertainty-Aware Trajectory Optimization Framework to Enhance Visual Tool Tracking for Robust Surgical Automation
- Real-to-Sim Registration of Deformable Soft Tissue with Position-Based Dynamics for Surgical Robot Autonomy
- SurRoL: An Open-source Reinforcement Learning Centered and dVRK Compatible Platform for Surgical Robot Learning
- Reinforced Grounded Action Transformation for Sim-to-Real Transfer
- CAD2RL: Real Single-Image Flight Without a Single Real Image
- Control of a Quadrotor With Reinforcement Learning
- Extracting Strong Policies for Robotics Tasks from Zero-Order Trajectory Optimizers
- Transcriptome Sequencing Unveils a Molecular-Stratification-Predicting Prognosis of Sarcoma Associated with Lipid Metabolism
- Inhibition of hyaluronic acid degradation pathway suppresses glioma progression by inducing apoptosis and cell cycle arrest