VoxelSage, a multimodal system that combines CT analysis with preoperative liver-resection planning, reported lower simulated time and blood loss than a deterministic controller in a test of 256 unseen scenes. The result is an engineering finding, not evidence that the system improves surgery: the authors say their results show system integration and simulator-level performance, not clinical efficacy or safety.
On those scenes, the learned, shielded controller averaged 33.388 minutes, compared with 34.274 minutes for the deterministic baseline. Its mean simulated blood loss was 183.852 mL, compared with 300.847 mL for the baseline. That amounted to a 0.886-minute, or 2.59%, time difference and a 116.995 mL, or 38.89%, blood-loss difference.
The time result was consistent in the paired comparison: C4, the learned shielded controller, was faster than C0 in 252 scenes, tied in three and lost in one. The C4-minus-C0 difference was -0.886 minutes, with a reported 95% confidence interval from -0.947 to -0.826 minutes and a reported Cohen's d_z of -1.77.
What the simulator actually tested
The planning study used disjoint, frozen data splits: 448 scenes for policy training, 64 for internal development, 64 for tuning, 128 for validation, 128 for a one-shot test and 128 for stress testing. The primary endpoint was the paired difference in simulated time between C4 and C0. A separate C4-versus-C2 comparison tested learning specificity, C3 served as a full-rollout teacher reference, and C5 was the version with exact verification removed. Confidence intervals were calculated from 10,000 paired bootstrap resamples.
That comparison also favored the learned controller against the specified non-learned comparator. C4 was 3.736 minutes faster than C2, with a reported 95% confidence interval of -4.094 to -3.389 minutes and a reported Cohen's d_z of -1.28.
Budget behavior was different when the verification shield was removed. In the nominal set, unshielded C5 exceeded its episode budget in 24 of 256 scenes, while C4 had no overruns. Across five sensitivity conditions with 128 scenes each, C4 completed every scene; C5 recorded 14, 10, 12, 11 and 8 overruns.
The system behind the numbers
VoxelSage divides the workflow between two ports. Port A handles request interpretation and chooses the appropriate skill, while Port B performs CT and mask computation. The system is designed for two- and three-dimensional CT visualization, liver-tumor analysis and preoperative resection planning.
The physical-coordinate stage was tested on synthetic phantoms. Across 24 phantoms, all 186 numerical comparisons and all 30 boundary-condition checks passed, with numerical comparisons required to fall within 0.01 in the corresponding metric unit. The check supports the reported computation tolerance, not a claim of clinical efficacy or safety.
On three public cases, the online evaluation ran all eight built-in Port B skills. The functional matrix recorded 66 checks across 24 unique skill-case paths, and all passed. Median cold preparation took 43.3 seconds; cached reuse was reported as 0.0 seconds at the stated one-decimal precision.
Useful checks, limited reach
Port A's task tests covered 154 tasks: lesion existence, lesion counting, liver volume, total lesion volume and maximum three-dimensional lesion diameter. Under the study's experimental agreement criteria, it was correct on 44 of 44 lesion-existence tasks, 27 of 27 total-lesion-volume tasks and 25 of 25 diameter tasks. Performance was lower on counting, at 30 of 34, and liver volume, at 22 of 24.
In a separate multi-case test, Port A identified the case with the larger liver volume correctly in 59 of 63 comparisons. It returned no invented measurements for nonexistent cases in all 137 queries, and 130 of 137 repeated queries needed no additional tool calls. The median response time changed from 8.16 seconds to 2.56 seconds, a reported 68.6% reduction, although batch answer accuracy was 122 of 137.
Those figures describe what the workflow did under its test criteria. They do not establish clinical efficacy or safety, a boundary the authors state explicitly.
The clinical reach is also narrow. The image-analysis interface used liver, tumor and vessel masks, but did not encode whether a tumor was primary liver cancer or a metastasis. Results from the colorectal-liver-metastasis cases therefore do not establish generalization to primary liver cancer or other tumor origins.
The work is an arXiv preprint identified in the supplied document as arXiv:2609.37648v1 and dated 29 September 2026. It leaves clinical efficacy, safety and broader tumor-origin generalization unresolved.
Paper data and sources
Original title: VoxelSage: Tool-Augmented 3D CT Analysis and Simulator-Shielded Sequential Resection Planning for Liver Tumors
Authors: Binghong Qian, Xuanhe Liu, Yifan Xing et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-09-29
DOI: Not available
Original paper · Full text