An uncertainty-guided human-review system reported better tumor-boundary scores than an automated baseline in an independent 362-case MRI evaluation, according to a new preprint. For the whole tumor, mean Dice, a measure of volumetric overlap, was 0.914 with the system versus 0.891 for the baseline. Mean HD95, a measure of boundary distance, was 4.76 millimetres versus 5.82 millimetres.
The largest reported change came in the tumor core. Mean HD95 fell from 17.96 millimetres for the baseline to 14.83 millimetres for the UG-HITL result, while mean Dice rose from 0.286 to 0.356. The authors described that as a 24.5% relative Dice improvement and a reduction of more than 3.1 millimetres in the boundary-distance measure. These figures come from the computational evaluation.
The review burden was uneven
The system was designed to connect automated segmentation with targeted boundary refinement and route structural failures for review. Its SA:V complexity rule set a review budget ranging from 5% to 25% of the total tumor volume. A structural collapse, defined as a base tumor volume of less than 1,000 voxels or more than four disconnected primary masses, sent the case to 100% manual contouring. The rule therefore included a full-review route for structural failures.
Across the UTSW cohort, the flagged workload averaged 20.2%, with a median of 11.3% and a maximum of 196.5%. The maximum can exceed 100% because workload was calculated against ground-truth volume for hallucinated cases. The structural tripwire activated in approximately 10% of cases: 36 of the 362 cases were routed to full manual review.
Several choices shaped the result
The baseline was a 3D full-resolution nnU-Net. To estimate uncertainty, native test-time augmentation used four geometric forward passes. The evaluation used Dice and HD95 to assess overlap and boundary distance, then applied a deterministic Human Oracle that overwrote flagged predicted voxels with ground-truth labels. The oracle was a theoretical upper bound, representing a best-case correction for the simulated review pathway.
On curated in-distribution data, the EDL configuration recorded a Dice score of 0.990, an HD95 of 0.62 millimetres and a review workload of 1.58%. A five-model ensemble, another uncertainty engine, recorded a Dice score of 0.931, below the automated baseline's 0.943. These were side-by-side computational comparisons, and no confidence intervals or formal significance tests were reported.
One additional metric was harder to read. For the ET category, mean Dice was 0.528 for the baseline and 0.577 for UG-HITL with TTA, but the median was 1.000 in both groups. The paper attributes that gap to subjects with zero ET voxels receiving perfect empty-overlap scores.
The safety net still has a blind spot
The paper describes a failure mode the tripwire can miss: a high-confidence, contiguous contralateral hallucination. In plain terms, the segmentation can contain a connected false region on the opposite side that looks internally coherent enough to bypass the structural check. The maximum post-refinement whole-tumor HD95 reported was 109.15 millimetres, showing how a system focused on flagged uncertainty can still miss a confident but wrong structure.
In a 50-case UTSW ablation subset, topological and anomaly filtering recorded whole-tumor Dice of 0.896 and HD95 of 6.71 millimetres. The full gradient-snapped pipeline kept whole-tumor Dice at 0.896 but had HD95 of 6.85 millimetres. For tumor core, the corresponding Dice, HD95 and workload figures were 0.369, 17.66 millimetres and 18.92% with the filtering configuration, compared with 0.368, 17.76 millimetres and 18.89% for the full pipeline.
Speed remains a practical constraint. Full TTA inference took approximately 14 to 18 seconds per patient volume, while CPU triage added only milliseconds. The paper identifies that latency as a bottleneck for real-time intraoperative deployment. The document is arXiv preprint version 1, dated 1 October 2026, and the authors report no funding and no competing interests.
The datasets are reported as publicly available from their corresponding repositories. The wider message is narrower than a clinical promise: the evaluation reported better boundaries and routed approximately 10% of cases to full review, while its deterministic oracle was a theoretical upper bound and its structural tripwire had a documented blind spot.
Paper data and sources
Original title: Uncertainty-Guided Handshake: Efficient Human-in-the-Loop Refinement for Surgical-Grade Glioma Segmentation
Authors: Samuel Hart, Ahmad Yahya, Ahmed Karam Eldaly
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-10-01
DOI: Not available
Original paper · Full text