PaperlayerView exact revision on arXiv

Source-linked AI summary

Occupancy Network-Guided Autonomous Robotic Partial Nephrectomy

Ethan Kilmer, Pit Henrich, Jiawei Ge, Paul M. Scheikl, Laura Connolly, Soum D. Lokeshwar, Joseph Chen, Justin D. Opfermann, Kaitlyn Kumar, Lauren Shepard, Ahmed Ghazi, Nirmish Singla, Richard J. Cha, Kevin Cleary, Franziska Mathis-Ullrich, Axel Krieger

arXiv:2609.16186v1cs.ROcs.CV

TL;DR

Tumor resection must achieve negative margins while balancing complete excision against preserving healthy tissue. This paper uses simulation-trained occupancy networks to infer anatomy and guide supervised robotic partial nephrectomy. In patient-derived hydrogel phantoms, the robot completed eight resections with negative margins and a 1.61 ± 0.48 mm mean absolute margin error.

  • Problem

    Tumor resection must remove the cancer while balancing the risk of incomplete removal against unnecessary loss of healthy surrounding tissue.

    Source: §I · p. 1

  • Method

    The system repeatedly infers 3-D tumor, margin, and kidney anatomy from partial point clouds using a simulation-trained occupancy network for supervised open partial nephrectomy.

    Source: C. Task Description · p. 3, §X · p. 18

  • Results

    Eight autonomous resections comprising 77 electrosurgical cuts achieved negative margins in all procedures, with a 1.61 ± 0.48 mm mean absolute margin error.

    Source: D. Contributions · p. 3

  • Takeaways & Limitations

    Accurate autonomous tumor resections were achieved in patient-derived hydrogel kidney phantoms under an open partial nephrectomy setting.

    Source: C. Task Description · p. 3, §X · p. 18

    The study could not statistically compare resection methods reliably because spatially correlated margin samples and small effective patient-level sample sizes limited the analyses.

    Source: C. Limitations · p. 17

Abstract

from arXiv · show

Autonomous soft-tissue cancer surgery has been limited to interventions on organ surfaces, because current systems cannot perceive and adapt to anatomy once it deforms or is cut. We introduce the first vision-guided autonomous system capable of performing complete tumor resections for partial nephrectomy. Our system integrates conditional occupancy networks, trained entirely in a physics-based simulation, that infer full 3-D anatomy (tumor, margin tissue, and kidney) from single-view partial point clouds. These occupancy networks maintain intraoperative tracking even as tissue is cut and deformed, enabling adaptive planning and execution. The surgical platform combines a depth camera for capturing surface point clouds, dual robotic arms for electrosurgical cutting and vacuum-based tissue manipulation, and an autonomous control strategy for tumor resection. In patient-derived hydrogel phantoms under an open partial nephrectomy setting, the robot performed eight consecutive autonomous tumor resections comprising 77 electrosurgical cuts, with all cuts achieving negative surgical margins and 1.61 $\pm$ 0.48 mm mean absolute margin error. This work demonstrates, for the first time, a foundation for supervised autonomous closed-loop, imaging-driven, margin-negative tumor removal in phantoms.

I. INTRODUCTION

Precise tumor resection must achieve a negative surgical margin without unnecessarily removing healthy tissue, yet deep dissection can hide the tumor and undermine image-based guidance. Earlier autonomous soft-tissue tumor resections did not address this full challenge; this work presents a camera-guided robotic system for resection.

I. INTRODUCTION

  • The system combines a depth camera, an occupancy network for 3-D kidney structure, and autonomous cutting with electrosurgical and vacuum-gripper tools.The network provides updated structural information as dissection progresses.Source: Figure 1 · p. 1
  • A negative surgical margin indicates no cancer cells at the specimen edge, but too narrow a margin risks incomplete resection while too wide sacrifices healthy tissue.An inadvertent cut into the tumor can cause a positive margin that may remain undetected until pathology results return.Source: §I · pp. 1–2
  • During deep margin dissection, tissue deformation and cutting undermine preoperative image mapping, and surrounding margin tissue can completely occlude the tumor.Near-infrared fluorescence signals do not provide reliable subsurface cut-plane distance, while ultrasound is not always reacquired during dissection.Source: §I · p. 2

A. Autonomous Robotic Tumor Resection

  • Only three prior approaches removed 3-D tumor-like masses, and each bypassed deep margin dissection or did not target both tumor and margin tissue.They addressed peeling a tumor without a margin, cutting a margin without a tumor, or palliative resection where incomplete removal was acceptable.Source: A. Autonomous Robotic Tumor Resection · p. 2

B. Kidney Cancer and Partial Nephrectomy

Partial nephrectomy removes a tumor and surrounding tissue while preserving the kidney, but deep dissection lacks continuous subsurface guidance. In a supervised open-surgery phantom setting, the proposed workflow repeatedly infers anatomy and replans, achieving eight autonomous resections with negative margins.

B. Kidney Cancer and Partial Nephrectomy

  • Partial nephrectomy removes the tumor and some surrounding tissue while preserving renal function; a negative surgical margin is associated with improved postoperative function and reduced recurrence risk.Reported positive surgical margin and overall complication rates remain 5.5% and 18.7%, respectively.Source: B. Kidney Cancer and Partial Nephrectomy · p. 2
  • Surgeons typically scan before resection for surface marking, then rely on visualization and mental mapping because continuous subsurface guidance during deep dissection remains unavailable.Reacquiring ultrasound can prolong the operation, disrupt workflow, and fail across air gaps created by cuts.Source: B. Kidney Cancer and Partial Nephrectomy · pp. 2–3

C. Task Description

  • The study targets supervised open partial nephrectomy on patient-derived hydrogel phantoms, aiming for a 6 mm margin tissue size.It combines patient-specific adaptive imaging, model-based workflow control, and human checkpoints for dynamically updated dissection plans.Source: C. Task Description · p. 3, D. Contributions · p. 3

D. Contributions

  • The workflow repeatedly infers complete 3-D anatomy from single-view partial point clouds through cutting, deformation, and tumor occlusion, supporting online replanning.It is trained in physics-based simulation and achieved 2.13 ± 1.04 mm tumor centroid error across 16 resection states.Source: D. Contributions · p. 3
  • Eight consecutive autonomous resections comprised 77 electrosurgical cuts, achieved negative margins in all procedures, and had a 1.61 ± 0.48 mm mean absolute margin error.Mean signed margin error was +0.41 ± 1.04 mm, with positive values indicating larger-than-planned clearance.Source: D. Contributions · p. 3

II. RELATED WORK

Prior autonomous tumor-resection work has addressed surface cuts or simplified removal tasks, while soft-tissue registration methods face challenges from deformation and alignment. The paper also situates its phantom approach among tumor models used in surgical research.

A. Autonomous Tumor Resection

  • Earlier autonomous tumor-resection studies included tumor peeling, surface incisions, and airway resections, but none addressed imaging through deformation, cutting, and occlusion during deep margin dissection.Only three prior studies achieved complete tumor-mass removal, and they did not address intraoperative imaging under those conditions.Source: A. Autonomous Tumor Resection · p. 3, Figure 3 · p. 4

B. Soft Tissue Registration

  • Rigid registration methods such as ICP can fail with increasing deformation, while CPD and learned deformation fields can be sensitive to initialization or require rigid pre-alignment.Rigidly aligning a rigid prior to a deformable object is described as inherently ill-posed.Source: B. Soft Tissue Registration · pp. 3–4
  • Conditioning an occupancy network directly on an observed point cloud produces a deformed 3-D object without requiring an initial rigid alignment.Prior occupancy approaches learned object shape from single-view observations; LUDO used FEM-based simulation-generated training data for deformable-object guidance.Source: B. Soft Tissue Registration · p. 4

C. Tumor Model

  • Tumor-model research spans patient tumors, animal models, artificial mimics, and synthetic phantoms, with synthetic phantoms supporting reproducible surgical evaluation without ethical constraints.Examples include patient-CT-based hydrogel kidney phantoms and silicone renal tumor phantoms.Source: C. Tumor Model · p. 4

III. SYSTEM OVERVIEW AND TASK FORMULATION

The Multi-shot workflow repeatedly infers patient-specific anatomy from exposed kidney surfaces, updates the dissection plan, and executes cuts with a supervised dual-arm robot. It uses phase-specific guidance for surface incision and deep dissection, including when the tumor becomes occluded.

III. SYSTEM OVERVIEW AND TASK FORMULATION

  • The workflow repeatedly captures the exposed kidney surface, infers full 3-D anatomy, updates the dissection plan, and executes it with a supervised dual-arm platform.Occupancy-guided inference supplies adaptive scene understanding; the robot performs cuts and tissue manipulation.Source: §III · p. 4

A. Multi-shot Task Decomposition

  • In progressive deep dissection, the tumor can become occluded as margin tissue, deformation, and cuts prevent direct fitting of the preoperative model to the scene.Grasping and cutting must coordinate through iterative retraction and cutting.Source: A. Multi-shot Task Decomposition · pp. 4–5

B. Anatomical Representation for Guidance

  • The guidance representation comprises kidney (K), circumferential cut (C), margin tissue (M), and tumor (T), with PDD using K, M, and T after C is removed.C guides the initial incision; M defines the intended resection shell, and T marks the boundary of oncologically unsafe cuts.Source: B. Anatomical Representation for… · p. 5

A. Intraoperative Imaging and Guidance

  • Partial kidney surface point clouds lack the 3-D structure needed for safe guidance, while registration becomes challenging after cutting and deformation.The paper introduces a cut-aware conditional occupancy framework to predict intact and progressive anatomy from observations, using LUDO.Source: A. Intraoperative Imaging and Guidance · p. 5
  • A conditional occupancy network adapts its 3-D representation to an observed kidney surface point cloud, reflecting cuts present in that observation.Figure 4 compares images after progressive depth cuts, the depth-sensor observation, and the network output.Source: A. Intraoperative Imaging and Guidance · p. 5, Figure 4 · p. 5
  • Separate conditional occupancy networks guide CSI surface incision and PDD deep dissection, using patient-specific meshes and paired simulated observations and occupancy ground truth for training.A unified perception representation is used across both phases; CSI could also use traditional registration because no kidney deformation is assumed.Source: A. Intraoperative Imaging and Guidance · p. 5

B. Occupancy Network: 3-D Mesh Creation

  • Patient CT segmentations are converted to meshes for the kidney, tumor, circumferential cut, and margin tissue, with Boolean operations constructing the guidance structures.The circumferential cut uses a hollow cylinder offset 5.5 mm and 7.7 mm from the tumor and extending 10 mm into the kidney; the margin tissue is formed by enlarging the tumor by 6 mm along vertex normals.Source: B. Occupancy Network: 3-D Mesh Creation · pp. 5–6

C. Occupancy Network: Physics-based Simulation

The PDD training pipeline simulates post-cut, deformed anatomy and varies cut surfaces and observations to train occupancy networks on progressive resection states. It pairs simulated observations with labeled occupancy data for learning.

C. Occupancy Network: Physics-based Simulation

  • PDD training simulates progressive cutting and deformation of K, M, and T, omitting C because the circumferential cut has already been removed.The simulation produces post-cut states rather than continuously simulating the cutting action; the paper contrasts this with mesh- and particle-based methods –.Source: C. Occupancy Network: Physics-based… · p. 6
  • A virtual cutting plane generates kidney cuts at a desired resection percentage, with random ±5° plane rotations modeling real-world cutting-path variability.The tumor remains attached to margin tissue through shared vertices and faces.Source: C. Occupancy Network: Physics-based… · p. 6
  • Layered Perlin-noise augmentation varies already-cut surfaces by up to 3 mm at low frequency and 0–1 mm at high frequency, leaving uncut points unchanged.The augmentation is intended to expose the network to irregular cuts while preserving structural information about the optimal margin.Source: C. Occupancy Network: Physics-based… · p. 6, Figure 6 · p. 6
  • Training observations are generated by masking kidney regions with a cut-off plane and vacuum gripper, then randomizing kidney orientation and cut-off-plane position.The virtual scene also randomizes the kidney rotation and plane location to produce the observation point cloud.Source: Figure 7 · p. 6
  • SOFA physics simulation deforms the cut meshes while a vacuum gripper retracts the tumor, saving successive K, M, and T states for training.The process generates 10,072 triplets per patient using a tetrahedral corotational FEM force field.Source: C. Occupancy Network: Physics-based… · p. 6
  • The PDD training loop conditions the occupancy network on a simulated observation and updates it using loss between predicted labels and labeled ground-truth points.Ground-truth occupancy point clouds are paired with surface point clouds resembling CSI- and PDD-phase camera observations.Source: Figure 8 · p. 7

D. Occupancy Network: Training Models

The occupancy models are trained on simulated phase-specific observations and integrated into a supervised finite-state workflow. The robot repeatedly captures anatomy, plans cuts, and executes coordinated cutting and retraction through CSI and PDD.

D. Occupancy Network: Training Models

  • PDD training observations come from randomly cut and deformed SOFA meshes, with anatomy masked, kidney orientation varied by up to ±35° per axis, and the cut-off plane shifted by up to ±25 mm.Depth images are captured at 256×256 resolution, and Improved Sort Sample provides occupancy ground truth.Source: D. Occupancy Network: Training Models · p. 7
  • The CSI and PDD datasets contain 32,000 and 50,000 training examples, respectively, each pairing an observation with an occupancy point cloud.The training configuration and cost are summarized in Table I, with shared parameters across patient models within each phase.Source: D. Occupancy Network: Training Models · p. 7, Table I · p. 7

V. AUTONOMOUS RESECTION WORKFLOW

  • The workflow uses an FSM planner with human approval of state transitions and the option to re-plan before execution.The FSM sequence covers initialization, CSI, repeated PDD, and tumor removal; Figure 9 shows representative states and images.Source: §V · p. 7, Figure 9 · p. 8
  • The planner avoids fully end-to-end learning because PSMs may be unobservable during surgery and oncologically irreversible, enabling preexecution intervention on unsafe plans.The implementation runs a multithreaded ROS 2 executor at 10 Hz.Source: §V · p. 7
  • The robot captures and segments the kidney point cloud, then uses the phase-specific occupancy model to label anatomy for planning.A supervisor supplies the initial SAM 2 point prompt, reused for subsequent captures.Source: §V · pp. 7–8
  • CSI paths are generated from incision-rim points for supervisor approval, then executed with the cutting arm; vacuum grasping establishes initial tissue tension.Rejected plans return to scene capture for replanning; the gripper is positioned and lifted to establish tension before deep dissection.Source: §V · pp. 8–9
  • PDD partitions margin points into 3 mm-spaced slices and coordinates each planned cut with incremental vacuum retraction before recapturing anatomy for the next iteration.The loop ends after tumor-width-based iterations or supervisor confirmation of complete removal; the retraction profile adjusts lateral, vertical, and tilt movements across iterations.Source: §V · p. 9

VI. ROBOTIC PLATFORM

The robotic platform provides the camera-guided, dual-arm hardware and control substrate for the perception and planning pipeline. Its arms divide electrosurgical cutting and vacuum-based tissue manipulation.

VI. ROBOTIC PLATFORM

  • The platform combines a camera and electrosurgical cutting arm with a separate vacuum-based grasping arm, mounted on a hydraulic table.The paper describes the platform as the execution substrate rather than its primary methodological contribution.Source: §VI · p. 9, A. Hardware Architecture · p. 9

A. Hardware Architecture

  • The cutting arm uses a 6-DoF UR10e, RGB-D camera, laparoscopic extension, and needle electrode.The camera and cutting arm are described as a combined subsystem.Source: A. Hardware Architecture · p. 9
  • The grasping arm uses a 6-DoF UR5 manipulator, electric vacuum gripper, moisture trap, laparoscopic extension, and interchangeable bellows pads.The pads can be swapped for different tumor sizes; the paper states vacuum grasping reduces tissue damage relative to forceps.Source: A. Hardware Architecture · p. 9
  • The auxiliary setup includes an electrosurgical unit, grounding pad, smoke evacuator, and air purifier for electrosurgical odor mitigation.The authors selected FDA-approved equipment or products from manufacturers with existing FDA approvals for future clinical translation.Source: A. Hardware Architecture · p. 9

B. Software Architecture

The platform coordinates calibrated camera and robotic-arm frames, supports synchronized task- and joint-space motion, and uses layered interlocks requiring supervisor approval before cutting. Experiments use patient-derived, electrically conductive hydrogel phantoms molded from CT data.

B. Software Architecture

  • Camera and manipulator frames are aligned to a shared checkerboard-defined world frame for coordinated multi-arm, camera-guided control.The system supports task-space trajectories with synchronized Cartesian position and quaternion orientation, and joint-space trajectories with synchronized linear interpolation.Source: B. Software Architecture · p. 9

C. Safety Architecture

  • Four safety layers combine hardware controls, a real-time halt, speed limits, and supervisor approval of every CSI and PDD plan before cutting.The cutting speed provides approximately a 3 s supervisor response window across a 6 mm margin.Source: C. Safety Architecture · p. 10

VII. EXPERIMENTS

  • The study evaluates occupancy-guided anatomy inference against CT references and assesses autonomous tumor removal in an open-field setup mimicking OPN access.The software uses ROS 2 Humble on Ubuntu 22.04, with the electrosurgical unit manually enabled because it is not programmable.Source: B. Software Architecture · pp. 9–10, §VII · p. 10

A. Phantom Fabrication and Setup

B. Experimental Design: 3-D Anatomy Inference

The inference experiments compare occupancy networks with registration and rigid-motion baselines using paired CT and depth observations across resection states. Evaluation measures tumor localization and 6 mm target-margin identification against CT-derived ground truth.

B. Experimental Design: 3-D Anatomy Inference

  • Paired CT volumes and depth point clouds are collected at five resection states across four patient models, yielding 20 registered data pairs for evaluating inference against CT ground truth.The collection includes CSI and four PDD states, from 0% to 75% PDD.Source: B. Experimental Design: 3-D Anatomy… · p. 11, Figure 13 · p. 11
  • CSI occupancy inference is compared with ICP, CPD, and V2S, while PDD inference is compared with a rigid-motion baseline and a model trained without cut-surface randomization.The registration methods receive approximate manual rigid alignment; the rigid baseline applies vacuum-gripper motion to the initial tumor and margin predictions.Source: B. Experimental Design: 3-D Anatomy… · p. 11

C. Evaluation Metrics: 3-D Anatomy Inference

  • Evaluation measures 6 mm target-margin error and tumor localization by centroid error, Hausdorff 95% distance, and IoU against registered CT segmentation.Margin error is measured from the CT tumor because the phantom has no distinguishable margin contrast; PDD performance differences are tested across 16 scenes, while CSI is not statistically analyzed because of low sample size.Source: C. Evaluation Metrics: 3-D Anatomy… · p. 12

D. Experimental Design: Robotic Resection

The robotic-resection experiments compare iterative multi-shot guidance with a single-shot rigid-assumption baseline and a surgeon, using a 6 mm target for the robotic groups. The section also reports CSI and PDD anatomy-inference results against registration and rigid-motion baselines.

D. Experimental Design: Robotic Resection

  • The study uses four phantoms per patient model, a 30 W electrosurgical setting, and vacuum pads sized for the tumor models.Multi-shot guidance uses iterative point-cloud capture and inference for a 6 mm margin; the single-shot baseline uses one initial observation and assumes rigid tumor and margin motion.Source: D. Experimental Design: Robotic… · p. 12

E. Evaluation Metrics: Robotic Resection

  • Robotic resection evaluation records NSM success, margin size and consistency, target-margin error, removed kidney volume, and procedure duration; the surgeon group has no target-margin error metric.Postoperative CT-derived margin surfaces are sampled to calculate distances to the tumor; COV is standard deviation divided by mean margin, expressed as a percentage.Source: E. Evaluation Metrics: Robotic Resection · p. 12, Figure 14 · p. 13

A. 3-D Anatomy Inference

  • CSI occupancy inference achieved the lowest mean centroid error, 3.08 ± 0.94 mm, versus ICP, CPD, and V2S; it also had the lowest Hausdorff 95% distance and highest IoU.The CSI values were 3.60 ± 0.63 mm for Hausdorff 95% and 0.56 ± 0.07 for IoU.Source: A. 3-D Anatomy Inference · p. 13
  • PDD occupancy inference had 1.97 ± 1.55 mm absolute margin-identification error and 2.13 ± 1.04 mm centroid error, compared with 2.13 ± 1.62 mm and 6.29 ± 2.28 mm for the rigid baseline.Differences were significant for absolute margin error, centroid error, Hausdorff 95%, and IoU, but not signed margin error.Source: A. 3-D Anatomy Inference · p. 13
  • Without cut-surface randomization, PDD localization error was 2.35 ± 0.85 mm, Hausdorff 95% was 3.39 ± 1.21 mm, and IoU was 0.65 ± 0.06.These results are reported for the model trained without surface randomization.Source: A. 3-D Anatomy Inference · p. 13

B. Robotic Resection

In eight multi-shot procedures, every resection achieved a negative surgical margin, while the single-shot baseline failed in all four procedures. The inference analysis describes advantages over rigid tracking alongside limitations related to unmodeled deformation and margin metrics.

B. Robotic Resection

  • Eight multi-shot resections all achieved NSMs, with 1.63 ± 1.13 mm average absolute margin error against the 6 mm target.The average margin size was 6.63 ± 1.88 mm, with 2.17 ± 0.92 mm average minimum margin size and 28.31% COV.Source: B. Robotic Resection · p. 13, Table VI · p. 15
  • Single-shot resection success was 0% across four experiments, with three tumor punctures and one trial aborted for unacceptable tool proximity.Its average absolute margin error was 2.38 ± 1.53 mm; one emergency halt occurred, while multi-shot had none.Source: B. Robotic Resection · pp. 13, 15
  • The surgeon achieved NSMs in all four procedures, with 58.82% COV and 0.57 ± 0.16 mm average minimum margin size.Average healthy kidney volume removed was 5245.75 mm3, and average procedure duration was 3 min and 46 s.Source: B. Robotic Resection · p. 15

A. 3-D Anatomy Inference

  • The PDD occupancy network significantly outperformed the rigid-assumption method on absolute margin error, centroid error, Hausdorff distance, and IoU, but not signed margin error.The signed measure is described as an indicator of over- or under-estimating the margin rather than overall accuracy.Source: A. 3-D Anatomy Inference · p. 15
  • Vacuum retraction drew tumors into the cup, a deformation absent from simulation training that likely contributed to localization-metric variation.The effect was clearest for the narrow Patient 4 tumor, which deformed substantially within the cup and had higher-than-average Hausdorff distances.Source: Figure 15 · p. 14, A. 3-D Anatomy Inference · p. 15
  • Margin-identification error is measured against the 6 mm target because the desired margin is not distinguishable in CT, and the metric does not account for margin deformation.The network tracks the deforming margin from the pre-resection state.Source: A. 3-D Anatomy Inference · p. 15
  • Rigid tracking misses tumor deformation that occurs upon vacuum activation before gripper motion, whereas occupancy inference maintains localization during deformation and tissue removal.Standard deformable registration assumes fixed topology, which is unsuitable when cutting directly changes the tumor–margin region.Source: A. 3-D Anatomy Inference · pp. 15–16
  • Surface randomization produced modest average improvements across localization metrics, alongside slightly increased standard deviations.Tumor localization error changed from 2.35 ± 0.85 mm to 2.13 ± 1.04 mm, while IoU changed from 0.65 ± 0.06 to 0.66 ± 0.08.Source: A. 3-D Anatomy Inference · p. 16

B. Robotic Resection Performance Analysis

Multi-shot used iterative scene feedback to resect tumors with more consistent margins than the surgeon, while the surgeon preserved more healthy kidney tissue. Robot dexterity, margin selection, observation transitions, and grasping or segmentation events constrained performance.

B. Robotic Resection Performance Analysis

  • Multi-shot successfully resected tumors, while Single-shot failed because it underestimated stretched margins without vision-based feedback and physics-based scene understanding.A patient-specific margin offset could have allowed Single-shot to succeed without positive surgical margins, but would require preoperative prediction.Source: B. Robotic Resection Performance… · p. 16
  • Multi-shot and the surgeon achieved negative surgical margins in all tumors; Multi-shot margins were more consistent, while the surgeon preserved more healthy kidney tissue.Deep dissection challenged the surgeon when healthy tissue occluded the tumor boundary; Multi-shot continued updating scene understanding during those stages.Source: B. Robotic Resection Performance… · pp. 16–17
  • Patient 4’s tumor location made robot access and trajectory planning difficult, producing surface incisions that deviated from simulated training data.Improved robot positioning or wristed tool degrees of freedom could mitigate these access issues.Source: B. Robotic Resection Performance… · p. 17
  • The 6 mm margin selection led Multi-shot to remove more kidney tissue than the surgeon, and repeated posture changes between imaging and cutting extended procedure duration.Future work will optimize margin size and explore vision guidance without switching arm postures.Source: B. Robotic Resection Performance… · p. 17
  • Segmentation artifacts sometimes prompted supervisor re-planning, and three vacuum-grasp failures with Patient 2 required manual reengagement before autonomous resection continued.The small tumor apex reduced the bellows pad’s sealing area; future work includes more robust segmentation and automated failure recovery.Source: B. Robotic Resection Performance… · p. 17
  • A supervisor halted one Single-shot trial after unacceptable proximity between the electrosurgical tool tip and vacuum pad; no collision occurred.This left three Single-shot trials reported although four had been planned.Source: B. Robotic Resection Performance… · p. 17

C. Limitations

The study demonstrates autonomous tumor resection in patient-derived hydrogel kidney phantoms, while identifying limits in retraction generalizability, simulation coverage, phantom realism, and statistical comparison. Translation to in vivo or minimally invasive procedures requires further validation or substantial system changes.

C. Limitations

  • The model-based progressive vacuum retraction trajectory may not generalize to other tumor resection procedures.Future work proposes online vision-guided planning and learning-based tissue manipulation for adaptive retraction.Source: C. Limitations · p. 17
  • The occupancy network cannot account for scene appearances unanticipated by simulation, limiting direct extension to other resections without new simulations and anticipated deformation scenarios.The study’s predictable state-based robot system made simulation feasible in this setting.Source: C. Limitations · p. 17
  • The non-perfused hydrogel phantom does not bleed, so the study does not evaluate autonomy for bleeding or hemostatic control.Future work proposes perfused phantoms and improved FSM-based autonomy for unexpected events.Source: C. Limitations · p. 17
  • Tumor size and shape may differ between preoperative imaging and surgery, which the current occupancy network cannot account for; it also requires per-case retraining.Proposed directions include multimodal guidance, subject-specific adaptation, robotic palpation, and generalized or zero-shot tracking conditioned on preoperative data.Source: C. Limitations · p. 17
  • Spatially correlated margin samples and dependent resections limit statistical comparisons, while small effective sample sizes leave some tests underpowered.The paired comparisons had effective sample sizes of N = 3 versus Single-shot and N = 4 versus Surgeon; consecutive cuts are also dependent.Source: C. Limitations · p. 17

D. Future Clinical Translation

  • Clinical translation toward in vivo open partial nephrectomy remains distinct from minimally invasive translation, and the former requires validation in perfused in vivo models.The proposed in vivo porcine work would address absent perfusion and anatomical mismatch; phantom properties also vary across subjects.Source: D. Future Clinical Translation · pp. 17–18
  • Minimally invasive partial nephrectomy would require new depth estimation and vision pipelines, wristed tools with remote-center control, and extended collision planning.Peritumoral fat, surrounding organs, tool occlusion, and electrosurgical smoke challenge segmentation and occupancy inference.Source: D. Future Clinical Translation · p. 18

X. CONCLUSION

  • The system achieved accurate autonomous tumor resections in patient-derived hydrogel phantoms under an open partial nephrectomy setting.The conclusion describes a Multi-shot workflow using occupancy-based anatomy inference and a hybrid learning and model-based autonomy stack.Source: §X · p. 18

APPENDIX

The appendix supplements the vision-pipeline analysis with runtime reporting and evaluations of segmentation artifacts, observation angle, simulation-to-real transfer, and occlusion.

A. Timings

  • Perception latencies are reported jointly for the CSI and PDD models, with the hardware specified; Table VIII is titled as covering optimal margin identification and tumor localization error.The supplied table captions do not provide the corresponding cell values.Source: A. Timings · p. 18, Table VII · p. 18, Table VIII · p. 18

C. Segmentation Artifacts

  • 13.16 ± 11.29 mm average tumor-centroid error with un-eroded SAM 2 segmentation contrasts with 2.13 ± 1.04 mm in Table III, highlighting the importance of segmentation and point-cloud filtering.Binary mask erosion suppresses depth-image boundary artifacts that otherwise introduce outlier points.Source: C. Segmentation Artifacts · pp. 18–19, Figure 17 · p. 18

D. Observation Angle

  • 1.92 ± 0.54 mm average centroid error across varied simulated observation angles compares with 2.14 ± 0.36 mm for the single-angle real-world evaluation.The angle evaluation used 400 simulated inferences across four PDD states for Patient 2.Source: D. Observation Angle · p. 19

E. Simulated Occupancy Evaluation

  • 0.92 ± 0.42 mm average centroid error on unseen simulation states compares with 2.14 ± 0.36 mm in the real-world evaluation, an increase of 1.22 mm when translating to the real world.The test used 400 new Patient 2 simulation states absent from training.Source: E. Simulated Occupancy Evaluation · p. 19

F. Observation Occlusion

  • Predicted tumor-centroid standard deviation was ±0.49 mm without occlusion, ±1.85 mm with surgical instruments, ±0.78 mm with blood, and ±1.51 mm with both.The evaluation used four point-cloud captures of a partially resected hydrogel kidney phantom.Source: F. Observation Occlusion · p. 19, Figure 18 · p. 19

Related papers

Loading 2609.16186v1…