PaperlayerView exact revision on arXiv

Source-linked AI summary

Uncertainty-Guided Handshake: Efficient Human-in-the-Loop Refinement for Surgical-Grade Glioma Segmentation

Samuel Hart, Ahmad Yahya, Ahmed Karam Eldaly

arXiv:2610.01452v1cs.CV

TL;DR

Automated glioma segmentation can conceal localized catastrophic failures, while conventional interactive tools require clinicians to search manually for errors. UG-HITL combines TTA uncertainty with topological triage to route failures and prioritize localized refinement. On an independent clinical cohort, standard-case review was capped at 25% of the target tumor volume, with total baseline failures routed to full manual review.

  • Problem

    Automated segmentation can conceal localized catastrophic boundary failures in out-of-distribution clinical data, while conventional interactive tools make clinicians search manually for errors.

    Source: Introduction · p. 1

  • Method

    UG-HITL uses TTA uncertainty and hierarchical post-hoc topological rules to isolate structural anomalies and create localized geometric proposals for targeted clinician refinement.

    Source: Introduction · p. 2

  • Results

    On an independent cohort of 362 cases, standard-case review was capped at 25% of target tumor volume, while total baseline failures were routed to full manual review.

    Source: Introduction · p. 3

  • Takeaways & Limitations

    The framework offers a Pareto-efficient pathway for targeted verification and refinement on the evaluated out-of-distribution clinical cohort.

    Source: Discussion · p. 10

    The triage engine cannot generate a segmentation for an anatomical sub-region the baseline feature extractor entirely misses, as reflected in bounded Tumor Core Dice on UTSW.

    Source: Discussion · p. 10

Abstract

from arXiv · show

While state-of-the-art automated models for medical image segmentation achieve high mean performance, they frequently suffer from localized, catastrophic failures that preclude safe clinical deployment, particularly in neuro-oncology. Interactive segmentation frameworks mitigate this by incorporating human oversight, but traditionally impose prohibitive cognitive and temporal workloads by requiring clinicians to manually search for errors. In this project, we present an efficient, Hybrid Structural-Aleatoric Human-in-the-Loop framework for glioma segmentation that bridges the gap between automated baseline performance and surgical-grade precision, achieving sub-2.0 mm HD95 on curated benchmarks while providing safety-net routing for structural failures across real-world clinical data. By extracting voxel-wise Test-Time Augmentation (TTA) uncertainty and applying hierarchical topological filtering, our method proactively isolates high-risk structural anomalies. We comprehensively evaluated our approach on a challenging out-of-distribution clinical stress-test cohort (N = 362). Operating under a simulated Human Oracle, the framework improved the Whole Tumor (WT) Dice score from 0.891 to 0.914 and reduced the 95th percentile Hausdorff Distance (HD95) from 5.82 mm to 4.76 mm. Critically for surgical safety, the system rescued severe boundary failures in the Tumor Core, reducing mean HD95 from 17.96 mm to 14.83 mm (improving absolute TC Dice to 0.356). These spatial rescues were achieved while demanding a median interactive workload of just 11.3% of the target volume. Acknowledging this as a simulated upper bound lacking real-world cognitive friction, the framework nevertheless demonstrates a highly Pareto-efficient pathway for safely deploying clinical AI.

Introduction

Automated glioma segmentation can conceal catastrophic localized failures in out-of-distribution clinical data, while conventional interactive tools leave clinicians to search manually for errors. UG-HITL turns uncertainty into localized geometric proposals and routes severe structural failures for full review.

Introduction

  • OOD clinical data expose localized catastrophic boundary failures that global segmentation averages can conceal, including false-positive hallucinations and omitted lesions.The text identifies these worst-case topological failures as a barrier to safe autonomous deployment.Source: Introduction · pp. 1–2
  • Conventional interactive segmentation requires clinicians to search 3D scans slice by slice, imposing cognitive and temporal costs vulnerable to fatigue and oversight.The proposed alternative is active triage that guides attention to regions of peak spatial risk.Source: Introduction · p. 2
  • Existing uncertainty maps are mainly passive overlays, leaving an open gap in translating predictive variance into constrained, physically snapped geometric boundaries for rapid verification.The text also notes ensemble inference costs and uncertainty-calibration concerns under distribution shift.Source: Introduction · p. 2
  • UG-HITL applies TTA uncertainty and post-hoc hierarchical topological rules to isolate structural anomalies and construct gradient-snapped proposals for targeted refinement.Unlike topology-preserving training methods, the rules operate on frozen outputs without retraining or architectural modification.Source: Introduction · p. 2
  • The framework is presented as backbone-agnostic and uses spatial uncertainty to reduce the clinician’s visual search to a localized boundary mask, with behavior characterized across three uncertainty estimators.The estimators are MC Dropout, deep ensembles, and evidential uncertainty.Source: Introduction · p. 2
  • On controlled benchmarks, the framework targets surgical-grade precision defined as HD95 below 2.0 mm, while its external-cohort safety routing caps standard-case review at 25% and sends total baseline failures to full manual review.The review cap is a maximum for standard cases; complete baseline failures are routed to full review.Source: Introduction · pp. 2–3

Methods

The methods combine TTA-based uncertainty extraction with dynamic review budgeting and a decoupled triage pipeline. A deterministic Human Oracle estimates routing efficiency as an upper bound, while the framework is also evaluated on an independent out-of-distribution clinical cohort.

Datasets and Clinical Cohorts

  • The framework uses BraTS 2023 adult glioma data for training and initial evaluation, then tests resilience on an independent UTSW out-of-distribution cohort of 362 cases.The BraTS holdout validation cohort contains 187 cases and uses seeded k-fold partitioning to prevent data leakage.Source: Datasets and Clinical Cohorts · p. 3

Baseline Architecture and Uncertainty Extraction

  • A 3D full-resolution nnU-Net uses four geometric TTA passes, and aligned outputs provide mean expected probabilities and aleatoric uncertainty associated with tumor boundary ambiguity.The four passes are the base orientation and inversions along the sagittal, coronal, and axial planes.Source: Baseline Architecture and Uncertainty… · p. 3

Complexity-Aware Dynamic Budgeting

Algorithmic Triage and Morphological Filtering

  • Track A routes predictions with tumor volume below 1,000 voxels or more than four disconnected primary masses to 100% manual contouring, bypassing boundary refinement.This structural tripwire is designed to detect catastrophic structural collapse.Source: Algorithmic Triage and Morphological… · p. 3, Figure 1 · p. 4

Track B: Topological Triage

  • For predictions passing the tripwire, Track B first applies connected-component filtering to remove core clusters lacking an edema boundary at zero interaction cost.The pipeline also routes disjointed satellite clusters meeting its probability and size criteria for review.Source: Algorithmic Triage and Morphological… · pp. 3–4

Algorithmic Triage and Morphological Filtering

  • Figure 1 depicts inference and uncertainty generation, structural-failure routing, refinement of passing cases, and delivery of localized proposals for clinician review.Accepted proposals are merged into the final refined mask.Source: Figure 1 · p. 4

Track B: Topological Triage

  • Track B refinement uses an asymmetrical geodesic search band and T1ce gradients to set anatomical barriers, then scores candidates to seek false negatives and snap to tissue boundaries.The remaining voxel budget is allocated across sub-regions using their distinct surface-area-to-volume ratios.Source: Algorithmic Triage and Morphological… · pp. 3–4

Evaluation Metrics and Human Oracle Simulation

  • The evaluation measures Dice, HD95, and the percentage of target tumor volume flagged for review, with flagged predicted voxels replaced by ground-truth labels in a deterministic Human Oracle simulation.This isolates algorithmic routing efficiency but is a theoretical upper bound because real users face cognitive friction and inter-rater variability.Source: Evaluation Metrics and Human Oracle… · p. 4

Results

On the 362-case UTSW clinical cohort, UG-HITL improved Whole Tumor and Tumor Core boundary metrics under simulated human refinement while keeping median workload low through risk routing. Component analyses and examples show how filtering and targeted uncertainty address different failure types.

Results

Discussion

UG-HITL shifts clinicians from exhaustive error-hunting to localized verification, but its refinements remain bounded by baseline feature extraction and a simulated Human Oracle protocol. The discussion also identifies a residual blind spot for isolated, high-confidence contralateral hallucinations.

Discussion

  • 11.3% median interactive workload accompanied rescue of severe baseline hallucinations in the out-of-distribution UTSW cohort.The framework combines Test-Time Augmentation uncertainty with strict biological connectivity rules.Source: Discussion · p. 10
  • EDL’s uniform probability distribution assigned p = 0.75 cumulative foreground probability in empty background air, causing false-positive breakdown when bypassing argmax to rescue false negatives.The authors report pivoting to TTA after out-of-distribution clinical noise exposed this vulnerability.Source: Discussion · p. 10
  • The triage pipeline cannot create a segmentation for an anatomical sub-region the baseline feature extractor misses, bounding absolute Tumor Core Dice on the UTSW cohort.The authors identify self-supervised domain adaptation or foundational medical visual encoders as future ways to improve primary feature representations.Source: Discussion · p. 10
  • A single contiguous, high-confidence contralateral hallucination can pass the localized geometric tripwire, and the cohort’s maximum post-refinement WT HD95 was 109.15 mm.Proposed future checks include hemisphere symmetry or center-of-mass comparison with anatomical atlases.Source: Discussion · p. 10
  • The 11.3% median workload is an algorithmic upper bound under simulated perfect clinician compliance and agreement with ground truth, not a measured clinical workload.The authors note that inter-rater variability and cognitive friction could change actual interaction metrics.Source: Discussion · p. 10

Conclusions

Across 362 out-of-distribution clinical cases, the framework targeted severe boundary failures while keeping median interactive workload at 11.3% of target volume. The conclusion reports structural-tripwire coverage of dominant failure modes, alongside a residual contralateral-hallucination blind spot.

Conclusions

  • A > 3.1 mm absolute reduction in mean Tumor Core spatial error accompanied a median interactive workload of 11.3% of target volume across 362 out-of-distribution clinical cases.The conclusion characterizes the workload as securely capped and the boundary failures as catastrophic.Source: Conclusions · p. 11
  • The structural tripwire intercepted fragmentation and volume-collapse failures under real-world clinical noise without costly localized model retraining.The conclusion notes a residual blind spot to high-confidence contralateral hallucinations.Source: Conclusions · p. 11
  • The authors characterize the framework as a Pareto-efficient pathway for practical clinical-AI deployment in high-stakes workflows.The stated scope is surgical-grade neuro-oncology.Source: Conclusions · p. 11
Loading 2610.01452v1…