Pith. sign in

REVIEW 4 major objections 4 minor 48 references

SPA: Efficient User-Preference Alignment against Uncertainty in Medical Image Segmentation

T0 review · 4 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Four-option picks teach a segmentation model your preference

desk verdict A genuinely useful multiple-choice interaction scheme and broad evaluation, but the main preference-adaptation claim is not supported because training simulates a user who always picks the candidate closest to the model's own output. read the letter →

arxiv 2411.15513 v2 pith:MSWFZHHF submitted 2024-11-23 eess.IV cs.CV

Jiayuan Zhu , Junde Wu , Cheng Ouyang , Konstantinos Kamnitsas , J. Alison Noble This is my paper · ORCID
classification eess.IVcs.CV
keywords userpreferencealignmentmedicalimagesegmentationuncertaintymodelingmixtureofGaussiansinteractivemultiple-choiceinteractionmulti-raterannotations
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes SPA, a medical image segmentation framework that learns a clinician's preferred way to draw ambiguous boundaries from a small number of multiple-choice selections. Instead of asking for pixel-level clicks or boxes, SPA shows one recommended mask and four representative alternatives; the user chooses the one closest to their intent, and the model updates a latent mixture-of-Gaussians preference distribution before offering a new set. The authors claim that within a few iterations this converges to the user's personalized segmentation, and they report an average Dice score of 89.68% after three iterations across fundus, CT, and MRI tasks, beating deterministic, uncertainty-aware, and interactive baselines. A human study with five medical professionals reports markedly less time and fewer interactions than the interactive baseline MedSAM, and the authors state that a user can segment 35% more images with 39% fewer iterations.

What carries the argument

The load-bearing object is the preference distribution $p_{\theta}(z) = \sum_{m=1}^{M} \pi_m \mathcal{N}(z \mid \mu_m, \sigma_m^2)$, whose latent samples condition the segmentation decoder. A neural Preference Distribution Generator predicts the mixture parameters from the user's chosen candidate; K-means clustering of $N$ stochastic predictions yields the $K$ candidate masks shown to the user; and a SAM-style prompt encoder turns the selected mask into the feedback embedding. Together these convert a user's discrete choice into a changed distribution over segmentation outputs.

What would settle it

Run SPA with a scripted user who always picks the representative candidate least similar to the current aggregation; if Dice against that user's held-out annotations does not improve relative to a no-feedback baseline, the adaptation shown in the paper is an artifact of self-confirming simulation rather than genuine preference alignment.

Watch

Extended reading notes

Core claim

The central claim is that the variability in how clinicians segment ambiguous pixels can be captured by a parameterized preference distribution $p_{\theta}(z)$, a mixture of $M$ Gaussians, and that this distribution can be adapted at test time from simple preference feedback. SPA samples $N$ latent variables from $p_{\theta}(z)$, conditions the segmentation network on each, clusters the $N$ predictions into $K$ representative candidates, and asks the user to pick one; the chosen candidate is encoded and fed to a Preference Distribution Generator that re-estimates the mixture parameters. The appendix argues that with enough interactions the distribution converges to the single Gaussian $\mathcal{N}(\mu_u, \sigma_u^2)$ of the target user, and experiments show predictions moving toward included clinicians and away from excluded ones over iterations. On the paper's terms, the result is that uncertain boundary decisions become a multiple-choice task rather than a pixel-refinement task, and the model can align to a new user's preference in roughly three interactions.

Load-bearing premise

Training feeds the model its own closest candidate as the simulated user's choice, so the model may never learn to respond to a user who selects something genuinely different from what the model predicts.

Editorial extensions

If this is right

  • If the claim holds, interactive segmentation can be reduced to choosing among a few distinct candidates, sparing clinicians from pixel-level clicking.
  • Uncertainty-aware models gain a feedback channel, so the many-similar-predictions problem of conventional probabilistic segmentation disappears.
  • The same user's adapted distribution transfers to new images, so personalization carries across a session.
  • The reported gains imply fewer failure cases at fixed Dice thresholds, which matters when annotation time is constrained.
  • Any segmentation task with multi-rater annotations could be fitted to the same preference-alignment loop.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's training protocol simulates the user's choice as the candidate closest to the model's own aggregated prediction; since that choice never disagrees strongly with the model, the reported adaptation may understate how the system behaves for a user who picks an outlier candidate. A scripted adversarial-user test would separate learned preference adaptation from self-confirmation.
  • The multiple-choice interface suggests that other high-stakes annotation settings, such as pathology outlines or radiotherapy target volumes, could use the same preference-distribution update with a domain-specific candidate generator.
  • One testable extension is to let the model choose which $K$ candidates to show so that the selected option carries maximal information about the user's preference, rather than using unsupervised K-means centroids.
  • The convergence proof assumes the user's selections are i.i.d. from one Gaussian component; real users may switch preference mid-session, which would require the distribution to forget old evidence.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. SPA proposes a test-time preference-alignment framework for medical image segmentation. The method models annotator-specific segmentation preferences as a mixture-of-Gaussians latent distribution pθ(z), generates N stochastic predictions per iteration, aggregates them into a recommended mask, and presents K representative candidates via K-means clustering. The user selects one candidate; the selection is encoded and fed to a Preference Distribution Generator that updates pθ(z), and the process repeats. The authors claim that SPA adapts to individual users with fewer interactions than click/box-based interactive segmentation, with state-of-the-art Dice scores across REFUGE2, LIDC-IDRI, and QUBIQ tasks. The paper includes ablations, a human user study, an efficiency analysis, and a supplementary convergence argument for the preference distribution.

Significance. If the central claim were validated, SPA would offer a practically valuable interface: replacing pixel-level corrections with a small number of multiple-choice selections is a genuinely attractive way to handle inter-annotator variability in clinical segmentation. The paper has noteworthy strengths: public code, evaluation across seven tasks and three imaging modalities, a human user study, and ablations of the main components. However, the core preference-adaptation mechanism is not established by the evidence as presented. The training-time user simulation in Algorithm 1 is self-confirming, the supplementary convergence proof is for an idealized model rather than the implemented network, and the automated evaluation does not specify how test-time user choices are simulated. These issues are load-bearing for the paper's central claim that the model learns to shift its latent distribution toward a real user's preferred alternative.

major comments (4)
  1. [Algorithm 1 (training loop)] The simulated user is implemented as ru = arg min_{rk} ||rk - yfin_app||2. This is self-confirming: in every training interaction the feedback is the representative candidate closest to the model's own aggregated output, so the Preference Distribution Generator never observes a selection that disagrees with the current prediction. The only supervision tied to the multi-rater target y is the segmentation loss LCE on yapp; the interaction embedding e_u = E_I(r_u) carries no information about a user preference different from the model's current behavior. As a result, the training signal for the central claim, namely shifting pθ(z) toward a user's preferred alternative, is absent. The paper needs either a user simulation that selects the candidate closest to a held-out rater or rater-combination, or direct evidence (e.g., real clinician selections at test time) that the generator responds to non-confirming choices.
  2. [Supplementary Section 7 (Eqs. 2-7)] The convergence proof does not model the implemented system. It assumes the interactions ru^(j) are i.i.d. scalar samples from N(µu, σu^2), whereas in the paper ru is a K-means segmentation mask selected from K candidates and then encoded by a SAM prompt encoder; pθ(z) is updated by an amortized MLP, not by exact posterior computation. The proof establishes only that an ideal GMM posterior concentrates on the true component as J→∞, not that the trained Preference Distribution Generator and the actual feedback channel converge. This proof therefore cannot support the claim that the preference distribution pθ(z) can adapt to specific user preferences based on new interactions.
  3. [Section 4.3.1 / Table 1] Table 1 reports single Dice scores with no error bars or confidence intervals. Several test sets are very small (e.g., QUBIQ brain tumor has 4 testing cases and brain growth has 5), and several reported gaps between SPA and the next-best method are under a point (e.g., REFUGE2 3-Iter: 85.42 vs 83.08 for SAM, or 88.56 vs 84.10 in the first block). The SOTA claim in Section 4.3.1 is load-bearing and needs repeated-run statistics, paired significance tests, or per-case variability. Please also specify how many runs or random seeds the reported numbers are based on, especially given the stochastic sampling in the method.
  4. [Supplementary Section 8 and Section 4.3.3] The automated test protocol describes how ground-truth combinations are generated but never states how the simulated user's choice ru is made at test time. This matters because Table 2's alignment analysis (predictions moving toward included clinicians and away from excluded ones) is exactly the evidence needed to distinguish genuine preference adaptation from optimization toward the weighted ensemble used as the training target. Without specifying whether the user selection is oracle-based (closest to the consensus or held-out target) or self-confirming (closest to yfin_app), the alignment results in Section 4.3.3 are uninterpretable.
minor comments (4)
  1. [Section 3.3 and Algorithm 1] The notation '10.5' in the aggregation formulas (e.g., yapp = 10.5 (1/N Σ yn)) is ambiguous: it may be an indicator threshold at 0.5 or a scaling constant. Please typeset it explicitly (e.g., 1[mean ≥ 0.5]) and explain what operation is intended.
  2. [Table 1] The '1-Iter' and '3-Iter' columns are applied to all method categories, but deterministic and uncertainty-aware methods are not interactive. Please clarify how these columns were computed for non-interactive baselines, or use separate columns for them.
  3. [Supplementary Table 6] The header for Table 6 shows two numbers per method without labeled subcolumns; please add explicit 'Dice 75%' and 'Dice 84%' subheaders so the reader can parse the results.
  4. [Section 4.3.2 / Human User Study] The human user study reports only five participants and no statistical test, session-order randomization, or inter-annotator variability. A significance test or per-case breakdown would strengthen the efficiency claim, which is otherwise anecdotal.

Circularity Check

2 steps flagged · score 7.0 of 10

Algorithm 1 trains SPA on self-confirming feedback (ru = argmin ||rk − yfin_app||2), so the preference-adaptation claim is partly circular; the appendix proof assumes the conclusion it claims to derive.

  1. self definitional [Algorithm 1 (Training Process), 'Simulate User Selection for the Next Iteration' lines]
    "Get K Representative Candidates: {rk}K k=1 = KMeans({yfin n }N n=1); Select preferred representative candidate: ru = arg minrk ∥rk − yfin app∥2;"

    The paper's central claim is that pθ(z) adapts to a user's preference through interactions ru. But in the training loop the 'user' interaction is defined as the representative candidate closest to the model's own aggregate output yfin_app. The interaction embedding eu = EI(ru) that drives the Preference Distribution Generator is therefore a function of the model's current prediction, so it never presents a genuinely different annotator preference. The only supervision tied to the true multi-rater label y is the segmentation loss LCE on yapp; the selection signal is not conditioned on y or on any held-out rater.

  2. other [Supplementary Section 7 (Theoretical Proof), Eq. (2) and surrounding text]
    "Let D = {ru(j)}J j=1 represent J interactions from user u, where each interaction ru(j) is generated i.i.d. from a specific component N (µu, σ2 u)."

    The proof's premise already contains the conclusion: it assumes that interactions are i.i.d. samples drawn from the user's own Gaussian component N(µu, σu^2). The KL-divergence argument then only shows that a GMM posterior can identify the component from which the samples were drawn — a standard property that has no dependence on SPA's architecture. In the implemented system, ru is not an i.i.d. Gaussian sample; it is a K-means centroid of segmentation masks selected by the user (or, in training, chosen as argmin distance to yfin_app) and encoded by a SAM prompt encoder. The posterior parameters are also produced by a six-layer MLP, not by the closed-form GMM update of Eq. (7).

full rationale

The paper's novelty is that a few discrete user choices can align the model with an unseen user's preference through the Preference Distribution Generator. The training procedure in Algorithm 1, however, defines the simulated user's choice ru as the candidate closest to the model's own current aggregate prediction yfin_app. That makes the interaction signal self-confirming: it never supplies a contradictory preference, so the generator has no direct training signal for the operation the central claim depends on — moving pθ(z) toward a user who selects a candidate disagreeing with the current prediction. The only label-driven signal is the segmentation loss on yapp, which improves Dice but does not validate preference alignment per se. The supplementary convergence proof does not repair this gap because it assumes interactions are i.i.d. Gaussian samples from the user's component, whereas the actual interactions are K-means segmentation-mask centroids encoded by a SAM prompt encoder and processed by a learned MLP generator. The automated evaluation section does not state how the user's choice is simulated for the quantitative alignment results, leaving open whether those results use an oracle selection or the self-confirming rule. The human user study is genuine external evidence of efficiency and reduced interaction time, but it does not directly measure convergence to a distinct preference beyond subjective satisfaction. Self-citations (MRNet, MSA, SAM) are not load-bearing here. Overall, the core adaptation claim is partially circular: the training feedback channel by construction cannot exhibit the divergent-preference behavior that the method claims to learn, while the theoretical support assumes that very behavior.

Assumptions & free parameters 5 free parameters · 4 assumptions · 3 invented entities

The central method rests on several hand-set hyperparameters (N=48 tuned on the evaluation dataset, K=4 chosen by convention, M=16 unexamined) and on modeling assumptions that are plausible but unproven: that user preferences are Gaussian mixtures, that clinicians are internally consistent, and that K-means centroids of predictions represent meaningful preference alternatives. The convergence claim is derived for an idealized interaction model, not the implemented neural pipeline.

free parameters (5)
  • N = 48 sampled predictions per iteration = 48
    Tuned on REFUGE2 (appendix: 24/36/48 gave Dice 83.05/84.70/86.22) and then used across all datasets, including REFUGE2 itself for final evaluation.
  • K = 4 representative segmentation candidates = 4
    Chosen ad hoc based on multiple-choice conventions in high-stakes tests, not derived from the problem or data.
  • M = 16 Gaussian mixture components = 16
    No justification is provided for this number; it is a hand-set hyperparameter.
  • MAX_USER_ITERATIONS = 6 = 6
    Arbitrary cap on interaction rounds used in training and efficiency analysis.
  • Ground-truth weighting A sampled uniformly from [0,1] in training = uniform [0,1]
    The stochastic combination of multi-rater labels is a modeling choice that defines the training target but is not derived from observed annotation behavior.
assumptions (4)
  • domain assumption Each user's decisions on uncertain pixels can be modeled as i.i.d. samples from a parameterized Gaussian mixture distribution pθ(z) in the latent space.
    Section 3.1: 'We assume that each user's decisions on uncertain/ambiguous pixels... can be modeled as i.i.d. samples drawn from a parameterized distribution.' This is the foundational modeling assumption.
  • ad hoc to paper With sufficient interactions, the preference distribution pθ(z) converges to N(µu, σu) for the interacting user.
    Supplementary Section 7 proves this only for an idealized setting where interactions are i.i.d. samples from the component; the actual implementation uses user selections of segmentation masks encoded by a network, so the proof's assumptions do not match the method.
  • domain assumption Individual clinicians have consistent annotation patterns, and their interaction behaviors are likewise consistent.
    Supplementary Section 6 motivates SPA from a four-clinician U-Net experiment showing within-clinician consistency and across-clinician variation. This extrapolates from annotation patterns to interactive behavior without direct evidence.
  • domain assumption K-means clustering of the N sampled predictions yields representative candidates that capture distinct, clinically meaningful user preferences.
    Section 3.4: 'Each cluster produces a representative segmentation candidate rk as the centroid of the cluster.' The paper does not establish that cluster centroids correspond to interpretable preference variants.
invented entities (3)
  • Preference distribution pθ(z), a mixture of Gaussians over latent conditioning signals
    purpose: Models user-specific segmentation preferences and generates diverse candidate segmentations at inference time.
    The latent distribution is an internal model construct; no external measurement validates its structure or that users correspond to Gaussian components in the latent space.
  • Preference Distribution Generator (MLP predicting GMM parameters from the interaction embedding)
    purpose: Amortized inference to update pθ(z) from user feedback at inference time.
    This network is trained on the paper's self-confirming simulated user selections, so its behavior under genuine user preference shifts is not independently evidenced.
  • Interaction embedding e_u produced by the SAM prompt encoder
    purpose: Encodes a user-selected representative candidate as feedback into the preference distribution generator.
    The embedding is a learned representation; no evidence outside the framework establishes that it captures preference-relevant information.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SPA: Efficient User-Preference Alignment against Uncertainty in Medical Image Segmentation." pith.science (2026). https://pith.science/paper/MSWFZHHF

@misc{pith2026241115513,
  author       = {Pith},
  title        = {Pith review of: SPA: Efficient User-Preference Alignment against Uncertainty in Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MSWFZHHF}},
  note         = {Machine review of arXiv:2411.15513}
}
read the original abstract

Medical image segmentation data inherently contain uncertainty. This can stem from both imperfect image quality and variability in labeling preferences on ambiguous pixels, which depend on annotator expertise and the clinical context of the annotations. For instance, a boundary pixel might be labeled as tumor in diagnosis to avoid under-estimation of severity, but as normal tissue in radiotherapy to prevent damage to sensitive structures. As segmentation preferences vary across downstream applications, it is often desirable for an image segmentation model to offer user-adaptable predictions rather than a fixed output. While prior uncertainty-aware and interactive methods offer adaptability, they are inefficient at test time: uncertainty-aware models require users to choose from numerous similar outputs, while interactive models demand significant user input through click or box prompts to refine segmentation. To address these challenges, we propose \textbf{SPA}, a new \textbf{S}egmentation \textbf{P}reference \textbf{A}lignment framework that efficiently adapts to diverse test-time preferences with minimal human interaction. By presenting users with a select few, distinct segmentation candidates that best capture uncertainties, it reduces the user workload to reach the preferred segmentation. To accommodate user preference, we introduce a probabilistic mechanism that leverages user feedback to adapt a model's segmentation preference. The proposed framework is evaluated on several medical image segmentation tasks: color fundus images, lung lesion and kidney CT scans, MRI scans of brain and prostate. SPA shows 1) a significant reduction in user time and effort compared to existing interactive segmentation approaches, 2) strong adaptability based on human feedback, and 3) state-of-the-art image segmentation performance across different imaging modalities and semantic labels.

Figures

Figures reproduced from arXiv: 2411.15513 by the authors.

Figure 1
Figure 1. (a) Existing uncertainty-aware models require users to choose from numerous similar-looking candidates, making the process [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overall framework of SPA. The inference process comprises of two steps: Preference-aware Segmentation and Preference Adaptation with Human Feedback. At iteration j, SPA takes the input image x, an interaction embedding e (j) u , and latent variables {zn (j) } N n=1 drawn from the preference distribution p (j) θ (z) to generate N segmentation predictions. These predictions are then combined into an aggregated prefere… view at source ↗
Figure 3
Figure 3. SPA Shows Superior Segmentation Visualization. Visual comparison of segmentation results with deterministic, uncertainty￾aware, and interactive models after six iterations. SPA provides better adaptability, particularly at boundary regions [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: SPA Demonstrates Extraordinary Efficiency. Efficiency analysis comparing the average number of iterations required to reach specific Dice Scores across interactive models. Models that failed to reach the target Dice Score within six iterations are assigned an iteration…
Figure 5
Figure 5. Figure 5: Representative Segmentation Candidates Converge Over Iterations. (a) Boxplot illustrating the similarity (measured by Dice Score) between representative segmentation candidates across multiple iterations. (b) Visualization showing the disagreement among four correction…
Figure 6
Figure 6. Figure 6: SPA Outperforms Other Interactive Models in Prediction Refinement. Visual comparison of predictions from SAM, MedSAM, MSA, SAM-U (V1, V2), and SPA models across six interaction iterations. The last column shows the difference between the final and initial predictions. …
Figure 7
Figure 7. Figure 7: SPA Demonstrates Visual Prediction Alignment with Clinician Annotations. The figure illustrates the differences between SPA’s optic cup segmentation predictions and individual clinicians’ annotations in the REFUGE2 dataset across multiple interaction iterations. Dark p…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 32 canonical work pages

  1. [1]

    Armato, Geoffrey McLennan, Luc Bidaut, Michael F

    Samuel G. Armato, Geoffrey McLennan, Luc Bidaut, Michael F. McNitt-Gray, Charles R. Meyer, Anthony P. Reeves, Binsheng Zhao, Denise R. Aberle, Claudia I. Hen- schke, Eric A. Hoffman, Ella A. Kazerooni, Heber MacMa- hon, Edwin J. R. van Beek, David Yankelevitz, Alberto M. Biancardi, Peyton H. Bland, Matthew S. Brown, Roger M. Engelmann, Gary E. Laderach, D...

    work page 2011

  2. [2]

    Fisher, Thomas R

    Byron Ballou, Gregory W. Fisher, Thomas R. Hakala, and Daniel L. Farkas. Tumor Detection and Visual- ization Using Cyanine Fluorochrome-Labeled Antibodies. Biotechnology Progress , 13(5):649–658, 1997. _eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1021/bp970088t. 1

    work page doi:10.1021/bp970088t 1997

  3. [3]

    Baumgartner, Kerem C

    Christian F. Baumgartner, Kerem C. Tezcan, Krishna Chai- tanya, Andreas M. Hötker, Urs J. Muehlematter, Khoschy Schawkat, Anton S. Becker, Olivio Donati, and Ender Konukoglu. PHiSeg: Capturing Uncertainty in Medical Image Segmentation. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2019, pages 119–127, Cham,

    work page 2019

  4. [4]

    Swin-Unet: Unet-like Pure Transformer for Medical Image Segmentation,

    Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xi- aopeng Zhang, Qi Tian, and Manning Wang. Swin-Unet: Unet-like Pure Transformer for Medical Image Segmentation,

    work page

  5. [5]

    Yuille, and Yuyin Zhou

    Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L. Yuille, and Yuyin Zhou. TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation, 2021. arXiv:2102.04306 [cs]. 1, 6

    arXiv 2021

  6. [6]

    FocalClick: Towards Practical Interactive Image Segmentation

    Xi Chen, Zhiyan Zhao, Yilei Zhang, Manni Duan, Donglian Qi, and Hengshuang Zhao. FocalClick: Towards Practical Interactive Image Segmentation, 2022. arXiv:2204.02574 [cs]. 1

    work page Pith review arXiv 2022

  7. [7]

    The Cancer Imaging Archive (TCIA): Maintaining and Operating a Public Information Repository

    Kenneth Clark, Bruce Vendt, Kirk Smith, John Freymann, Justin Kirby, Paul Koppel, Stephen Moore, Stanley Phillips, David Maffitt, Michael Pringle, Lawrence Tarbox, and Fred Prior. The Cancer Imaging Archive (TCIA): Maintaining and Operating a Public Information Repository. Journal of Digital Imaging, 26(6):1045–1057, 2013. 6

    work page 2013

  8. [8]

    A comparison between three-and four-option multiple choice questions

    Afsaneh Dehnad, Hayedeh Nasser, and Agha Fatemeh Hos- seini. A comparison between three-and four-option multiple choice questions. Procedia-Social and Behavioral Sciences, 98:398–403, 2014. 3

    work page 2014

Show all 48 references
  1. [9]

    SAM-U: Multi-box prompts triggered uncertainty estimation for reliable SAM in medical image, 2023

    Guoyao Deng, Ke Zou, Kai Ren, Meng Wang, Xuedong Yuan, Sancong Ying, and Huazhu Fu. SAM-U: Multi-box prompts triggered uncertainty estimation for reliable SAM in medical image, 2023. arXiv:2307.04973 [cs]. 3, 6, 5

    2023 arXiv

  2. [10]

    REFUGE2 Chal- lenge: A Treasure Trove for Multi-Dimension Analysis and Evaluation in Glaucoma Screening, 2022

    Huihui Fang, Fei Li, Junde Wu, Huazhu Fu, Xu Sun, Jaemin Son, Shuang Yu, Menglu Zhang, Chenglang Yuan, Cheng Bian, Baiying Lei, Benjian Zhao, Xinxing Xu, Shaohua Li, Francisco Fumero, José Sigut, Haidar Almubarak, Yakoub Bazi, Yuanhao Guo, Yating Zhou, Ujjwal Baid, Shubham Inn...

    2022 arXiv

  3. [11]

    Guan, Varun Gulshan, Andrew M

    Melody Y . Guan, Varun Gulshan, Andrew M. Dai, and Ge- offrey E. Hinton. Who Said What: Modeling Individual Labelers Improves Classification, 2018. arXiv:1703.08774 [cs]. 2, 6

    2018 arXiv

  4. [12]

    Masked Autoencoders Are Scal- able Vision Learners, 2021

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked Autoencoders Are Scal- able Vision Learners, 2021. arXiv:2111.06377 [cs]. 4

    2021 arXiv

  5. [13]

    Deep Learning Techniques for Medical Image Segmentation: Achievements and Challenges

    Mohammad Hesam Hesamian, Wenjing Jia, Xiangjian He, and Paul Kennedy. Deep Learning Techniques for Medical Image Segmentation: Achievements and Challenges. Journal of Digital Imaging, 32(4):582–596, 2019. 1

    2019

  6. [14]

    A review of uncertainty quantification in medical image analysis: Probabilistic and non-probabilistic methods

    Ling Huang, Su Ruan, Yucheng Xing, and Mengling Feng. A review of uncertainty quantification in medical image analysis: Probabilistic and non-probabilistic methods. Medical Image Analysis, 97:103223, 2024. 2 9

    2024

  7. [15]

    Improving Uncertainty Estimation in Convolutional Neural Networks Using Inter-rater Agreement

    Martin Holm Jensen, Dan Richter Jørgensen, Raluca Jalaboi, Mads Eiler Hansen, and Martin Aastrup Olsen. Improving Uncertainty Estimation in Convolutional Neural Networks Using Inter-rater Agreement. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2019, p...

    2019

  8. [16]

    Learning Calibrated Medical Image Segmentation via Multi- rater Agreement Modeling

    Wei Ji, Shuang Yu, Junde Wu, Kai Ma, Cheng Bian, Qi Bi, Jingjing Li, Hanruo Liu, Li Cheng, and Yefeng Zheng. Learning Calibrated Medical Image Segmentation via Multi- rater Agreement Modeling. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages...

    2021

  9. [17]

    What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?, 2017

    Alex Kendall and Yarin Gal. What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?, 2017. arXiv:1703.04977 [cs]. 1

    2017 arXiv

  10. [18]

    Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. Segment Anything, 2023. arXiv:2304.02643 [cs]. 1, 3, 4, 6, 5

    2023 arXiv

  11. [19]

    Aleatory or epis- temic? Does it matter? Structural Safety, 31(2):105–112,

    Armen Der Kiureghian and Ove Ditlevsen. Aleatory or epis- temic? Does it matter? Structural Safety, 31(2):105–112,

  12. [20]

    Simon A. A. Kohl, Bernardino Romera-Paredes, Clemens Meyer, Jeffrey De Fauw, Joseph R. Ledsam, Klaus H. Maier- Hein, S. M. Ali Eslami, Danilo Jimenez Rezende, and Olaf Ronneberger. A Probabilistic U-Net for Segmentation of Ambiguous Images, 2018. 2, 6

    2018

  13. [21]

    Simon A. A. Kohl, Bernardino Romera-Paredes, Klaus H. Maier-Hein, Danilo Jimenez Rezende, S. M. Ali Eslami, Push- meet Kohli, Andrew Zisserman, and Olaf Ronneberger. A Hierarchical Probabilistic U-Net for Modeling Multi-Scale Ambiguities, 2019. arXiv:1905.13077 [cs]. 6

    2019 arXiv

  14. [22]

    Simon A. A. Kohl, Bernardino Romera-Paredes, Clemens Meyer, Jeffrey De Fauw, Joseph R. Ledsam, Klaus H. Maier- Hein, S. M. Ali Eslami, Danilo Jimenez Rezende, and Olaf Ronneberger. A Probabilistic U-Net for Segmentation of Ambiguous Images, 2019. arXiv:1806.05034 [cs, stat]. 1

    2019 arXiv

  15. [23]

    Hongwei Bran Li, Fernando Navarro, Ivan Ezhov, Amirhos- sein Bayat, Dhritiman Das, Florian Kofler, Suprosanna Shit, Diana Waldmannstetter, Johannes C. Paetzold, Xi- aobin Hu, Benedikt Wiestler, Lucas Zimmer, Tamaz Ami- ranashvili, Chinmay Prabhakar, Christoph Berger, Jonas Wei...

    2024 arXiv

  16. [24]

    SimpleClick: Interactive Image Segmentation with Simple Vision Transformers, 2023

    Qin Liu, Zhenlin Xu, Gedas Bertasius, and Marc Niethammer. SimpleClick: Interactive Image Segmentation with Simple Vision Transformers, 2023. arXiv:2210.11006 [cs]. 1

    2023 arXiv

  17. [25]

    Segment Anything in Medical Images

    Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment Anything in Medical Images. Nature Communications, 15(1):654, 2024. arXiv:2304.12306 [cs, eess]. 3, 6, 5

    2024 arXiv

  18. [26]

    Prasanna, Helen B

    Pataje G.S. Prasanna, Helen B. Stone, Rosemary S. Wong, Jacek Capala, Eric J. Bernhard, Bhadrasain Vikram, and C. N. Coleman. Normal tissue protection for improving radiother- apy: Where are the Gaps? Translational cancer research, 1 (1):35–48, 2012. 1

    2012

  19. [27]

    U-Net: Convolutional Networks for Biomedical Image Segmentation,

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation,

  20. [28]

    Christian Rupprecht, Iro Laina, Robert DiPietro, Maximil- ian Baust, Federico Tombari, Nassir Navab, and Gregory D. Hager. Learning in an Uncertain World: Representing Ambi- guity Through Multiple Hypotheses, 2017. arXiv:1612.00197 [cs]. 1

    2017 arXiv

  21. [29]

    Erickson

    Tomas Sakinis, Fausto Milletari, Holger Roth, Panagiotis Ko- rfiatis, Petro Kostandy, Kenneth Philbrick, Zeynettin Akkus, Ziyue Xu, Daguang Xu, and Bradley J. Erickson. Interactive segmentation of medical images through fully convolutional neural networks, 2019. arXiv:1903.082...

    2019 arXiv

  22. [30]

    Petrov, and Anton Konushin

    Konstantin Sofiiuk, Ilia A. Petrov, and Anton Konushin. Re- viving Iterative Training with Mask Guidance for Interactive Segmentation, 2021. arXiv:2102.06583 [cs]. 1

    2021 arXiv

  23. [31]

    Shin, Suryakanth R

    Nima Tajbakhsh, Jae Y . Shin, Suryakanth R. Gurudu, R. Todd Hurst, Christopher B. Kendall, Michael B. Gotway, and Jian- ming Liang. Convolutional Neural Networks for Medical Image Analysis: Full Training or Fine Tuning? IEEE Transactions on Medical Imaging, 35(5):1299–1312, 20...

    2016 arXiv

  24. [32]

    Zuluaga, Rosalind Pratt, Premal A

    Guotai Wang, Wenqi Li, Maria A. Zuluaga, Rosalind Pratt, Premal A. Patel, Michael Aertsen, Tom Doel, Anna L. David, Jan Deprest, Sebastien Ourselin, and Tom Vercauteren. In- teractive Medical Image Segmentation using Deep Learning with Image-specific Fine-tuning. IEEE Transact...

    2018 arXiv

  25. [33]

    Medical SAM Adapter: Adapting Segment Anything Model for Medical Image Segmentation,

    Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical SAM Adapter: Adapting Segment Anything Model for Medical Image Segmentation,

  26. [34]

    Alexander

    Le Zhang, Ryutaro Tanno, Mou-Cheng Xu, Chen Jin, Joseph Jacob, Olga Ciccarelli, Frederik Barkhof, and Daniel C. Alexander. Disentangling Human Error from the Ground Truth in Segmentation of Medical Images, 2020. arXiv:2007.15963 [cs]. 2

    2020 arXiv

  27. [35]

    Medical SAM 2: Segment medical images as video via Segment Anything Model 2, 2024

    Jiayuan Zhu, Yunli Qi, and Junde Wu. Medical SAM 2: Segment medical images as video via Segment Anything Model 2, 2024. arXiv:2408.00874 [cs]. 3 10

    2024 arXiv

  28. [36]

    A Review of Uncertainty Estimation and its Application in Medical Imaging, 2023

    Ke Zou, Zhihao Chen, Xuedong Yuan, Xiaojing Shen, Meng Wang, and Huazhu Fu. A Review of Uncertainty Estimation and its Application in Medical Imaging, 2023. arXiv:2302.08119 [cs, eess]. 2 1 SPA: Efficient User-Preference Alignment against Uncertainty in Medical Image Segmentat...

    2023 arXiv

  29. [41]

    A preliminary experiment testing the impact of individual clinicians, conducted for the optic cup segmentation on REFUGE2 test set under U-Net’s structure with Dice Score (%)

    Motivation Details Table 4. A preliminary experiment testing the impact of individual clinicians, conducted for the optic cup segmentation on REFUGE2 test set under U-Net’s structure with Dice Score (%). The results indicate that the segmentation performance is consistent for ...

  30. [42]

    from a specific component N (µu, σ2 u)

    Theoretical Proof Let D = {ru(j)}J j=1 represent J interactions from user u, where each interaction ru(j) is generated i.i.d. from a specific component N (µu, σ2 u). The posterior probability that the samples (interactions) D comes from user (Gaussian component) p is given by:...

  31. [43]

    48 predictions is experimentally the best to balance model performance and computational cost

    Implementation Details To capture image uncertainty, we generate N = 48predic- tions by sampling from the preference distribution p(j) θ (z). 48 predictions is experimentally the best to balance model performance and computational cost. The Dice Score for generating 24, 36, 48...

  32. [44]

    We provide the failure rate statistics for REFUGE2 dataset reaching Dice 70% and 80%, for LIDC reaching Dice 60% and 70% in Table 9

    Efficiency Analysis on Different Interactive Models Details Models that failed to reach the target Dice Score within the limit were assigned an iteration count of ten. We provide the failure rate statistics for REFUGE2 dataset reaching Dice 70% and 80%, for LIDC reaching Dice ...

  33. [45]

    Overall Diff

    Representative Segmentation Candidate Similarity across Interactions In this section, we explore how the similarity between rep- resentative segmentation candidates evolves over multiple iterations of human interaction. Our multi-choice approach generates distinct segmentation...

  34. [46]

    Diff 1” refers to the improvement from Iteration 1 to Iteration 2, “Diff 2

    Prediction Change After Interactions In this section, we highlight the effectiveness of our SPA model’s multi-choice correction candidate interaction strat- egy, demonstrating how it outperforms other interactive mod- els in terms of prediction refinement. The quantitative res...

  35. [47]

    Five medical professionals, each with over five years of graduate-level expertise, participated in the study

    Human User Study Details In order to evaluate the efficiency of our SPA model com- pared to the previous interactive model, MedSAM, we con- ducted a detailed human evaluation study simulating real- world medical image segmentation workflows. Five medical professionals, each wi...

  36. [48]

    7 illustrates a visual comparison of the differences between SPA’s segmentation predictions and individual clin- icians’ annotations over six iterations

    Visualization for Prediction Alignment with Clinicians Fig. 7 illustrates a visual comparison of the differences between SPA’s segmentation predictions and individual clin- icians’ annotations over six iterations. The focus is on how the model’s predictions evolve with human i...

  37. [2015]

    arXiv:1505.04597 [cs]. 1, 6, 2

    arXiv

  38. [2019]

    Springer International Publishing. 1, 2

  39. [2021]

    arXiv:2105.05537 [cs, eess]. 6

    arXiv

  40. [2023]

    arXiv:2304.12620 [cs]. 3, 6, 5

    arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.