Abstract

Label-free chemical imaging holds significant promise for improving digital pathology workflows, but data acquisition speed remains a limiting factor. To address this gap, we propose an adaptive strategy—initially scan the low information (LI) content of the entire tissue quickly, identify regions with high aleatoric uncertainty (AU), and selectively re-image them at better quality to capture higher information (HI) details. The primary challenge lies in distinguishing between high-AU regions mitigable through HI imaging and those that are not. However, since existing uncertainty frameworks cannot separate such AU subcategories, we propose a fine-grained disentanglement method based on post-hoc latent space analysis to unmix resolvable from irresolvable high-AU regions. We apply our approach to streamline infrared spectroscopic imaging of breast tissues, achieving superior downstream segmentation performance. This marks the first study focused on fine-grained AU disentanglement within dynamic image spaces (LI-to-HI), with novel application to streamline histopathology. Code will be made public.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0694_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{OhJi_Finer_MICCAI2025,
        author = { Oh, Ji-Hun and Falahkheirkhah, Kianoush and Bhargava, Rohit},
        title = { { Finer Disentanglement of Aleatoric Uncertainty Can Accelerate Chemical Histopathology Imaging } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15972},
        month = {September},
        page = {202 -- 212}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper tackles the challenge of slow data acquisition in label-free chemical histopathology imaging, a key bottleneck for its clinical adoption. The authors propose an adaptive imaging strategy: perform a fast, low-information (LI) scan of the entire tissue, identify regions of high aleatoric uncertainty (AU), and then selectively re-image only those regions at high-information (HI) quality. The core contribution lies in addressing the challenge that not all high AU regions benefit from HI re-imaging. To this end, they introduce a method for finer-grained disentanglement of AU, separating “resolvable” uncertainty (which improves with HI data) from “irresolvable” uncertainty using post-hoc latent space analysis. This allows targeted HI acquisition to maximize information gain while minimizing imaging time. The approach is demonstrated on infrared spectroscopic imaging data of breast tissues for a segmentation task.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • Novel and practical application of uncertainty quantification (UQ) directed at accelerating the imaging process itself, moving beyond typical UQ uses like trustworthiness or active learning for labeling.  
    • It directly addresses one of the most important real-world problems in IR histopathology: the trade-off between imaging speed and data quality/information content.  
    • The core technical contribution is the conceptualization and implementation of fine-grained AU disentanglement (resolvable vs. irresolvable) specifically for a dynamic imaging setting.  
    • The proposed method for disentanglement is post-hoc and relies on latent space analysis (specifically Mahalanobis distance to prototypes), avoiding complex model retraining or architectural changes, which enhances its potential practicality. - The experimental validation demonstrates clear benefits over random sampling and a simpler “Max AU” baseline, particularly in improving calibration (ECE).
    • The paper includes ablation studies that test the sensitivity of the method to the quality of the underlying static UQ estimates and the fidelity of the latent space approximation, strengthening the results. Robustness checks with alternative distance metrics (KNN) and dimensionality reduction are also included.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • The figures have extremely bad resolution and quality. Even zoomed in, they are barely readable.
    • The experimental validation is based on a single infrared dataset for breast cancer, obtained from one cited source. This limits the ability to assess the generalizability of the findings to other tissue types, imaging modalities and data sources.  
    • While the application of fine-grained AU disentanglement to dynamic imaging acceleration is novel, the underlying UQ technique relies on established latent space distance methods. The paper could more clearly delineate the novelty compared to prior latent space UQ work.  
    • The reported improvements in the primary segmentation metric (F1 score) are modest, and the “Max AU” baseline actually performs slightly better in F1 score under the unconstrained budget.
    • The practical cost-benefit analysis could be expanded. While a cost function is defined and tested with a ~20% UAR requiring re-imaging, the implications for scenarios with much larger UAR ratios or different LI/HI cost differentials need further discussion.  
    • The paper primarily compares against random sampling and a naive max AU selection strategy. A discussion or comparison against alternative (non-UQ based) methods for accelerating chemical imaging or improving LI image analysis would strengthen the paper’s positioning.
  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
    • Definetely rework/reexport all figures to deliver readable quality.
    • Could you elaborate on the sensitivity of the method to the choice of thresholds \tau_{EU} and \tau_{AU}?
    • While the setting strategy is described, how robust is the final adaptive performance to variations in these thresholds?
    • The limited validation on a single dataset is a concern for generalizability. While new experiments aren’t possible, could you discuss which aspects of the approach are expected to generalize well (e.g., the latent space distance concept) and which might require tuning for different modalities or tissue types?
    • The paper shows strong gains in ECE and P(a,c) but less so in F1. Is there an interpretation for why targeting resolvable AU improves calibration more than raw segmentation accuracy compared to the Max AU baseline?
    • Consider adding a brief discussion contextualizing this UQ-based acceleration approach relative to other potential strategies for speeding up chemical imaging.
  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper presents a novel application of UQ to address a practical problem in chemical histopathology, introducing the interesting concept of fine-grained, dynamically resolvable AU. The methodology is technically sound, leveraging recent advances in latent space UQ, and the experiments demonstrate tangible benefits, particularly in calibration, supported by good ablation studies. However, the limited validation scope (single, super small dataset) raises questions about generalizability, and the improvement in the primary task metric is modest. The novelty lies more in the application and specific AU partitioning concept than in the base UQ technique itself. Acceptance could depend on the authors convincingly addressing generalizability concerns and further contextualizing the performance gains and practical scalability in their rebuttal.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    The main contribution of this work is a method for disentangling different types of aleatoric uncertainty in adaptive imaging. The authors propose a post-hoc latent space analysis to identify which high-uncertainty regions are likely to benefit from higher-quality imaging. According to the authors, this enables more efficient data acquisition and improved segmentation performance, particularly in the context of infrared spectroscopic imaging for histopathology. The method builds on the concepts of aleatoric and epistemic uncertainty, and is evaluated on a segmentation task involving breast cancer tissue, using a UNet model to extract the latent space.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The main idea of making the best use of limited resources in medical imaging is relevant, and their proposal seems to coherent with their problem. Their approach seems novel and the theoretical justification provided makes intuitive sense. As it is presented, it seems to be applicable to other use cases (types of images/tissues, architectures).

    The paper is reasonably well-written; the bibliography is extensive and also on point (not easy with only two pages). For instance, the concepts of “aleatoric” and “epistemic” uncertainty are not universally well-defined not accepted, yet they managed to put their version well into context before elaborating on their own method.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • Although the paper presents a reasonable case for the proposed method, evaluating it with at least one non-UNet architecture would strengthen the argument for its generalizability.

    • choice of notation: Section 3.1 is not immediately clear, and several choices make it hard to follow (EU vs UE, etc.)

    • going from the classification setup in Sec. 2 & 3 to segmentation in Sec.4 does not appear trivially evident: do your notions of uncertainty transfer seamlessly to segmentation? How? here more explanations are needed.

    • efficiency: you mention it in your abstract, but only go back to this concept explicitly at the very last page, with no context/interpretation for your statement

    Minor remarks:

    • page 2: “probability simplex” seems an unnecessary choice of vocabulary (uncommon for many members of the audience)
    • page 3: line 6: “which holds when…” : what holds? what form? unclear “static vs dynamic”: what is static and what is dynamic? references needed, these terms are too vague and not clear from context “identical samples”: in what sense? footnote: “…lipschitzness” grammatically sketchy and not clear, please use full English expression, or just drop the footnote
    • page 5: “performance upper-bounded” unclear: “it” what? why? “dataset”: how plausible is the downsampling? “implementations”: in general this paragraph is really not clear. “four models per domain …” this paragraph does not appear to be explicit enough if someone were to reproduce your experiment.
    • page 6: “uncertainty prototypes”: what is “prototype” in this context
    • page 7: “well-approximation of Eq. 4”: maybe more “ and (ii), a good approx…”?
  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    “weak accept”: although it already appears worthy of publication in its current form, it would benefit from:

    • more context to justify their application to a segmentation task (transition from Sections 2 & 3 to 4)
    • experiments with at least one different architecture (since it relies on a good latent space)
    • clearer figures, which are currently somewhat hard to read and not well contextualized
  • Reviewer confidence

    Somewhat confident (2)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    The authors present a novel method for uncertainty disentanglement with application to label-free digital histopathology imaging. To reduce the necessary imaging budget, the proposed method identifies regions on low-information image data that exhibit high disentangled aleatoric uncertainty to justify re-scanning to obtain high-information data. The method is well-described and extensively evaluated.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The paper is well-motivated and addresses an important problem in digital pathology.
    • Strong mathematical foundation.
    • Novel uncertainty disentanglement method tailored for digital pathology.
    • The paper is well-written.
    • The method is extensively evaluated and the results are properly discussed.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • This paper is already in a very good shape. The only major weakness is that the paper has a very high density and requires a lot of prior knowledge from the reader. It would benefit from more explanations or visualizations and I think it could have been submitted to a journal without a strict page limit.
    • The paper does not mention public code.
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
    • What is meant by “6K training pixels” in Section 4.1? Should it be 6K training images?
    • I disagree with the first sentence (§1): Histopathology is not directly a practice for treating cancer.
    • The comma-separated enumeration in sentence 2 (§1) has two “and” at the end.
  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I think that the paper is a strong submission to MICCAI, proposing a novel method, solving an important issue, and is well-evaluated. My only point of criticism is that the paper would benefit from more pages.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A




Author Feedback

We thank the reviewers for their positive feedback. All feedback will be incorporated; we address only key points below.

Shared (R1-R3):

  1. Reproducibility: Code will be made public.
  2. Figure quality: Visuals will be improved.

R1:

  1. Generalizability: Our method relies on (i) strong static EU/AU estimates and (ii) effective separation of UAI/UAR in latent space. For (i), we compute EU via latent space, leveraging its proven SOTA performance (Refs 34, 38, Vray ’24, Poceviciute ’25), and expect good generalization. For (ii), UAI/UAR are non-EU and should remain semantically clustered (thus, effective). The choice of distance function may require tuning though—e.g., KNN may outperform Mahalanobis for heterogeneous embeddings, but underperform with small samples due to noise.
  2. EU/AU thresholds: Setting them too low risks missing “good” data as it prunes UAR too aggressively; too high risks re-imaging “bad” data (e.g., high-EU, UAI data). This is ultimately a user design choice and our method is robust to reasonable values. However, as “reasonable” is subjective, we followed literature-guided protocols to methodologically set them. A sensitivity analysis would have helped, but space constraints limited this.
  3. Metric gains: Gains may seem modest due to being bounded by HI. When normalized against HI/LI bounds and random baselines (per Ref. 42, Malinin ’19, Amersfoort ’20), the gain is more evident: e.g., F1 of fine-grained AU under unconstrained budget is (55.8-53)/(60.5-53) ÷ (54.46-53)/(60.5-53)=1.92—a 92% improvement over random querying. We will add this context.
  4. Non-UQ baselines: A valid suggestion, but space limits led us to focus on proving our core claims—(i) limitations of monolithic AU and (ii) feasibility of AU partitioning—via max-AU comparison and ablation analysis. We will briefly acknowledge non-UQ strategies.
  5. Practicality: Our work included a “constrained” setting for cases where UAR exceeds the re-imaging budget. We agree that analyzing LI/HI cost differential role is interesting—e.g., lowering LI cost increases max query budgets and possibly net quality—but it adds analysis complexity beyond the page limit. We’ll briefly note this and explore it in future extension. However, we expect our method to remain broadly applicable, as it assumes no specific cost structure.
  6. On F1 gains and max-AU outperformance: Our method avoids consistently high-AU regions (UAI), like class borders. In LI, these borders are often inaccurate but well-calibrated; in HI, models may be more accurate, but overfit and hurt calibration. Max-AU prioritizes such data, likely explaining its better F1 but worse ECE and P(a,c). Arguably though, re-imaging such data is less valuable as they remain ambiguous (as per our AU thresholding).

R2: 6K refers to training pixels, as our task is segmentation.

R3:

  1. Notation and clarity: We will revise unclear terms, notations, pronouns, implementation details, and better connect efficiency to our method.
  2. Segmentation setup: As a dense per-pixel classification task, all concepts/notations transfer. We used segmentation due to dataset availability, but our method also applies to classification. This will be clarified.
  3. “Identical samples” = same tissue data
  4. Dataset downsampling (x4) is plausible since exact pixel-wise UQ resolution is not needed. The mild factor preserves tissue structure and does not affect results.
  5. Generalization across architecture: As noted (but omitted due to space), we also validated on Swin-Unet. While our method relies on well-structured latent spaces, this is supported in literature across CNNs and transformers (e.g., Ref. 38, Vray ’24, Poceviciute ’25). Our key contribution is the simple yet novel insight that latent spaces can also support AU estimation and partitioning—not just EU, as traditionally used.




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    This paper speeds up chemical histopathology imaging by first doing a quick scan with low-resolution imaging (LI) to find uncertain regions, then re-imaging only the parts in high-resolution imaging (HI) where more detail will actually help. All three reviewers appreciate the merits of the paper in terms of novelty—fine-grained AU disentanglement (resolvable vs. irresolvable), practical usage—the combination of algorithm and imaging itself, and extensive and relevant literature. I recommend paper acceptance based on the reviewers’ positive feedback, yet it is still advisable for the authors to address the remaining concerns of the reviewers, e.g., reproducibility and scalability of the algorithm, code availability, as well as discussion about the latent space.



back to top