acceptodds
Withdrawn from ICLR 2027

One Uncertainty Field, Two Scales: Gated Anatomy Refinement and Flip-Probability Diagnosis for Liver Tumors

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Liver-tumor segmentation and LI-RADS assessment are typically treated as independent modules: a segmentation network produces a mask, and diagnostic features are read off the mask with fixed thresholds, so the confidence of each voxel is discarded at the interface. We show that a single voxel-level uncertainty field , the Shannon entropy of the test-time augmentation posterior of an nnU-Net, operates at two scales. Spatially, weights a liver shape prior by local uncertainty, so the prior corrects the lesion boundary only where the network is unsure; this converts a prior that is harmful alone ( Dice) into a Dice improvement on WAW-TACE CT studies (; improve; paired Wilcoxon signed-rank ). Diagnostically, a mask threshold wandering through the uncertain band moves all phase means together, parameterizing a lesion-specific feature covariance ; under a three-feature linear surrogate of LI-RADS (APHE, washout, capsule) rather than the full clinical algorithm, this yields a closed-form Gaussian flip probability , the probability that measurement noise crosses the LI-RADS category boundary and reclassifies the lesion. On the lesions with a complete diagnostic readout, the propagated noise matches its boundary-perturbation Monte-Carlo reference (correlation ) but under-estimates the sampled mean flip rate -fold ( vs. ), whereas the fixed-noise baseline over-estimates it -fold () and correlates poorly (). We further show that refinement tends to help the least-stable tier of lesions — all five of the highest-flip-risk lesions stabilized, though this tier is small — while leaving already-confident decisions essentially unchanged. These results suggest that a single voxel-level uncertainty field can bridge segmentation and LI-RADS assessment, turning voxel confidence into both a refinement signal and a lesion-level flip probability.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
uncertainty quantification, uncertainty propagation, test-time augmentation, medical image segmentation, liver tumor segmentation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_K1MTOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Liver tumor assessment typically uses multiphase CT and features such as APHE, washout, and capsule appearance for LI-RADS classification. Motivated by this setting, the paper aims to use segmentation uncertainty both to improve segmentation and to assess diagnostic stability. The method applies nnU-Net to portal-venous-phase CT, averages the predicted probabilities obtained through test-time augmentation, and computes the Shannon entropy at each voxel to construct a shared uncertainty field. At the spatial level, this field controls the strength of liver shape and vessel suppression priors, directing corrections primarily toward regions where the model is uncertain. At the diagnostic level, the method varies the mask threshold across the uncertain boundary band to estimate a lesion-specific covariance that captures joint changes in diagnostic features across contrast phases. It then computes the probability of a diagnostic category flip under a Gaussian noise assumption.

Critical strengths
  1. Use of local uncertainty: The method uses local uncertainty to control where anatomical priors intervene, directing corrections primarily toward regions where the model is uncertain.
  2. Method interpretability: The authors model correlated errors in diagnostic features caused by changes in segmentation boundaries, construct a lesion-specific covariance, and relate it to diagnostic decision boundaries to compute category flip probabilities. This provides a clear, interpretable connection between segmentation uncertainty and diagnostic stability.
Critical weaknesses
  1. Limited generalizability: The experimental evaluation has limited coverage and a small sample size. Moreover, the benefits are concentrated primarily in over-segmented cases, with little improvement in under-segmented cases.
  2. Limited validity of the diagnostic flip probability: The score is primarily effective under boundary perturbations consistent with its modeling assumptions. It underperforms the fixed-noise baseline under independent perturbation mechanisms, relies on a simplified LI-RADS rule, and has not been validated against real-world clinical readings.
  3. Spatial refinement does not improve overall diagnostic stability.
Questions for authors

All three identified critical weaknesses are very serious.

Reviewer_azU7OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Automatic diagnosis of liver tumors is typically divided into two independent modules, where uncertainty in tumor segmentation has not been fully utilized. The authors argue that a valuable system should clarify how segmentation uncertainty affects diagnostic stability. Therefore, this paper explores using uncertainty in medical image segmentation simultaneously as a guide for correcting segmentation and as a tool for evaluating downstream diagnostic stability. Results on the WAW-TACE dataset show that the proposed anatomical optimization strategy improves the average Dice score from 0.653 to 0.683. In assessing diagnostic stability, the predicted flip probability outperforms a fixed noise baseline and exhibits a high correlation with boundary-perturbation-based Monte Carlo reference values. However, the strategy does not yield improvements on external LiTS/MSD datasets and performs poorly under other segmentation variation mechanisms. Furthermore, experiments indicate that although overall diagnostic stability is not enhanced, the optimization helps stabilize lesions with the highest flip risk.

Critical strengths
  1. The presentation and analysis of experimental results in this paper are objective and comprehensive. In addition to reporting results on the internal dataset, the authors validate the generalizability of their method using external datasets. They also provide a detailed analysis of the reasons behind suboptimal performance on external data, discussing the limitations and applicability conditions of the proposed approach.
  2. The paper systematically compares different segmentation perturbation mechanisms and uses this comparison to analyze the scope of application of the proposed uncertainty estimation method. The authors evaluate not only the flip probability estimation based on boundary perturbations but also compare it with MC Dropout, geometric test-time augmentation, and variations generated by different folds of models, illustrating differences in method performance across various perturbation mechanisms.
Critical weaknesses
  1. The results presented in the paper are insufficient to demonstrate the effectiveness of the proposed method. Although the authors objectively report results on both internal and external datasets, the effectiveness of the method remains questionable. It is unclear whether the performance improvement observed on the internal dataset is attributable to uncertainty itself or is related to the parameter settings and the choice of anatomical priors.
  2. The related work section does not sufficiently cover recent research. Most of the methods discussed in the paper are relatively old, with limited coverage of studies from the past two years. The current review of related work is insufficient to adequately support the novelty claims of the paper.
  3. The methodological implementation details and settings are not sufficiently described. The rationale for selecting the anatomical priors, as well as the modeling assumptions underlying the diagnostic flip-probability estimation and their applicability, requires further clarification.
  4. There are formatting issues with section and table titles, and some parts of the manuscript have inconsistent formatting. The title of Section 4.2 is incorrect, Section 4.3 contains formatting issues, and some table numbering is also incorrect. These issues affect the overall readability of the paper.
Questions for authors
  1. What are the substantive innovations of this paper compared with existing methods that combine uncertainty with anatomical priors? What new capabilities or advantages does this work offer for segmentation and diagnosis? Is the proposed method effective only on the internal dataset used in this study?
  2. Why were the signed distance field (SDF) and the vessel-suppression map selected as anatomical priors? Was the formulation in Equation (2) proposed in previous work or designed by the authors? What is the rationale behind the choice of the parameter 0.4?
  3. The paper states that each LI-RADS category is an intersection of multiple linear half-spaces and provides a closed-form expression for a single constraint. Can a category be defined by multiple simultaneous constraints? If so, how is the probability that all these constraints are satisfied computed?
Reviewer_qSVAOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper computes a voxel-level uncertainty field induced by a Shannon entropy defined over a probability obtained as the average of voxel segmentation probabilities (denoted by PTTAP_{TTA}) obtained by a nnUnet version, over four test-time augmentations of voxels. Then the paper proposes to use this voxel-level uncertainty field (named as u(x)\mathbf{u}(x)) as a single measurement guiding two different purposes for a downstream task of LI-RAD classification of liver tumors.

First purpose: A spatial confidence gate: αu(x)\alpha \mathbf{u}(x) is used as a multiplicative weight (in combination with a hyperparameter α\alpha) to gate an anatomical prior (defined using a signed-distance field Φ\Phi and a vessel supression map VV) which is then added to PTTAP_{TTA} to provide a better confidence metric PfinalP_{\text{final}} to the network's predictions.

Second purpose: A diagnostic LI-RAD rating flip probability: Using the posterior computed from the nnUnet, a three dimensional feature vector μ(t)\mu(t) comprising the arterial-phase hyperenhancement (APHE), washout and capsule features is computed using a sweep over the mask threshold tt. Using this feature vector set, a differential quantity Δμ\Delta \mu quantifying feature vector gaps between high-confidence regions and low-confidence regions is computed. Based on this differential quantity, a covariance Σseg(u)\Sigma_{\text{seg}}(u) is calculated. Along with a base noise, this covariance matrix characterizes a Gaussian whose mean is μ\mu. The feature vectors are then assumed to be samples from this Gaussian.

An important assumption is made that the LI-RAD grades are determined by linear separators. A quick computation yields a diagnostic flip probability to push a lesion class across boundaries. Three stability classes (LOW, MEDIUM, HIGH) are defined based on three arbitrary ranges of this flip probability.

Experiments are computed over WAW-TACE dataset volumes to evaluate the goodness of these spatial gates and diagnostic flip probabilities. Ablations on negative controls are provided for spatial gating. A Monte Carlo sampling based assessment using a sweep over the uncertain band is compared against the lesion stability of proposed flip probability. Further two other families are considered to measure flip rates against the proposed flip probability scheme.

Experiments of lesion stability across the LOW, MEDIUM and HIGH tiers show stability preservation on the LOW regime, where lesions are least stable. The caveat here is the low sample count (=5) in LOW regime. In both MEDIUM and HIGH regimes the stability results are only mildly impactful.

A validation conducted on an external cohort from LiTS dataset points towards limited transferability of both spatial gate effect and diagnostic score impact.

Critical strengths
  1. Derivative measures are computed using simply the segmentation results of a nnUnet version based on test time augmentations of the original CT volume, which serve two purposes: spatial gating and diagnostic probe into lesion grade (LI-RAD scale) stability.
  2. Extensive experiments are reported using CT volumes from a public dataset, with honest reports on failure modes.
Critical weaknesses
  1. Scope of the paper to ICLR is unclear: the paper simply proposes a statistical post-processing of the liver segmentation probability maps yielding derivate metrics for two different purposes: spatial gating and diagnostic probing of lesion LI-RAD grade swaps. The core argument that derived representations from liver CT volume segmentation maps serve as effective proxy LI-RAD features remains unjustified due to limited success with both in-domain and out-of-domain datasets (c.f. section 4.3, section 4.4., Table 2 and section 4.5). Is the paper trying to glorify an unsuccessful feature representation? Or is it trying hard to push the proposed uncertainty metrics as reasonable alternatives? I am afraid it's the former.
  2. LI-RAD grade measurement includes size and threshold growth as important features. It is unclear why these features were left out in the current study.
  3. Different LI-RAD scales reveal important characteristics of different aspects. Which of the LI-RAD scales are considered in the paper? What is the sample-wise distribution of the volumes across these grades? Some LI-RAD grades indicate benign nature of the lesion while some others indicate malignant nature, and some others reveal Hepatocellular Carcinoma status. Why should the proposed spatial gating and diagnostic probes serve uniformly across these different LI-RAD grades? A more detailed analysis of grade-wise flips, and grade-wise spatial gating results might give more insights into the proposal. This analysis however is missing.
  4. How APHE, non-peripheral washout and Capsule are measured in the current setup as LI-RAD features remains unclear. The discussion in Section 3.2 does not divulge details about explicit feature vector construction. In short: how are the three coordinates of ff computed using the mask threshold sweep procedure?
  5. For a 3D feature vector, wouldn't a non-linear classification boundary be more reasonable than a linear classifier (despite Figure 4)? While the linear assumption might provide a nice closed form expression for the stay probability, it might be a potential cause for the non-uniform stability results across the LOW, MEDIUM and HIGH regimes. This is not explored further in the paper.
  6. Empirical comparisons are limited to ablated designs of the same scheme. Meritorious comparisons with existing counterfactual techniques and other uncertain quantification techniques are possible but dismissed as irrelevant in the Introduction and related work sections. This is possibly the major setback to the current form of the paper, given the poor tier-wise results discussed in Section 4.4 and non-transferable mechanism utility on an out-of-domain dataset shown in Section 4.5.
  7. Perhaps the final devastating blow to the proposal in the paper is revealed in Section 4.5. Neither the spatial gating nor the diagnostic probe utility are transferable to an unseen-domain data. The paper pitifully succumbs to the serious issues discussed in section 4.5. What does the paper want to argue on the out-of-domain analysis made on the LiTS data in section 4.5? Since concerns of Ovadia et al. 2018 are already well-established, wouldn't a reasonable argument be to alleviate such concerns using a suitable proposal than to just conform to those concerns?

Writing and unspecified details:

a. The paper's writing appears quite incomprehensible to a human reader, even after multiple readings.

b. What are the four geometric transformations considered in T\mathcal{T} used in computing PTTAP_{TTA}?

c. What are the distance field Φ\Phi and vessel suppression map VV? Why are they the only indicators of anatomical prior? Citations missing for this use.

d. The discussion in section 3.2 is based on a posterior P. What is this P? Is it the predicted segmentation mask probability obtained from nnUnet without the test time augmentations? Also the discussion in section 3.2. involves a term called "phase". What is this?

e. Why is Δμ\Delta \mu based on the ranges >0.25> 0.25 and ≥0.75\geq 0.75? What happens if these ranges are changed?

f. Why are the tiers LOW, MEDIUM and HIGH based on the specific ranges of PflipP_{\text{flip}}? What happens if these ranges are changed?

Arbitrary experimental settings used in spatial gate and diagnostic probe construction:

i) Lines 213--214 indicate an additional experimental procedure in addition to eqn. (3). No ablations provided on this extra step.

ii) Arbitrary diagonals added to Σseg(u)\Sigma_{\text{seg}}(u) in eqn. (5) and in lines 247--248, without a discussion.

iii) σCT=\sigma_{CT} = 10HU? Why?

iv)Why is the base noise considered to be (5,8,6) HUs?

v) Though α=1\alpha=1 is emphasized in section 4.2, an elaborate discussion on the tiered stability results in section 4.4 uses different values of α\alpha to discuss different implications. Is α\alpha important or not?

vi) Filtering schemes in Section 4.1. and 4.5 seem to be inconsistent.

General comments:

While the overall thought behind the proposed uncertainty metrics is interesting, everything would have made more sense if the results from nnUnet on liver tumor segmentation was comprehensive. However as indicated in the paper, results on WAW-TACE and LiTS show around 0.56 or 0.68 Dice. In presence of such poor quantifications from a trained model, goodness measurement of an add-on metric like the ones proposed in the paper will be of limited value. While research to bring improvements in liver tumor segmentation are still ongoing, the paper could have chosen a similar problem on an organ where models show impressive behavior (e.g. kidney, brain, etc).

Further the primary hypothesis that LI-RAD can be proxied using just three features (with unspecified feature creation steps) should be rigorously tested. Including other LI-RAD characteristics should be looked into.

To make any proposed uncertainty metric useful for LI-RAD scales, differential analysis should be carried out across different severity scales of LI-RAD grades.

I would urge the authors to improve their work based on these comments.

Questions for authors

Authors can clarify only the most relevant points in my comments above in Weaknesses section.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.