List of Papers Browse by Subject Areas Author List
Abstract
Multimodal Large Models have significantly advanced automated breast ultrasound diagnosis. However, most existing frameworks utilize opaque, end-to-end paradigms prioritizing global statistical correlations over structured clinical reasoning. Consequently, these models remain susceptible to shortcut learning amid extreme real-world epidemiological imbalances, often bypassing rare but decisive malignant indicators for dominant benign patterns. To address this disconnect, we propose Latent-CURE, a novel diagnostic framework driven by asymmetric weighted chain-of-thought methodology grounded in latent space reasoning. Unlike traditional approaches, our framework constructs an implicit reasoning trajectory forcing the model to sequentially infer standardized BI-RADS morphological descriptors before converging on a final diagnosis. Furthermore, to combat the extreme scarcity of critical malignant features, we couple this architecture with a dual-asymmetric optimization strategy. By dynamically adjusting margins and weights, this strategy safeguards high-specificity malignant descriptors from being overshadowed by common benign priors. Comprehensive evaluations demonstrate that our knowledge-injected approach provides transparent clinical evidence while achieving robust, accurate diagnostic performance in imbalanced medical cohorts.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0748_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{ZhaWei_LatentCURE_MICCAI2026,
author = { Zhao, Weiyi AND Tan, Xiaoyu AND Gan, Lu AND Liu, Liang AND Qiu, Xihe},
title = { { Latent-CURE: Interpretable Breast Cancer Diagnosis via Dual-Asymmetric Chain-of-Thought } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16886},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors propose Latent-CURE, a novel diagnostic framework driven by asymmetric weighted chain-of-thought methodology grounded in latent space reasoning.
- 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.
Unlike traditional approaches, our framework constructs an implicit reasoning trajectory forcing the model to sequentially infer standardized BI-RADS morphological descriptors before converging on a final diagnosis. The authors evaluated their framework using an IRB-approved multicenter dataset of 666 breast ultrasound cases, partitioned 8:2 for training and testing, with all diagnostic labels strictly derived from the surgical histopathology gold standard. Table 2 shows quantitative diagnostic performance comparison across different model architectures, such as ResNet18 and various versions of Qwen. Based on the metrics used (accuracy, precision, specificity, and MCC), their proposed Latent-CURE comes out on top.
- 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 content is mostly good. Weaknesses are mostly related to presentation. The text in many of the figures is impossibly small to read. The last section “Discussion and Conclusion” lacks substance; it is too brief and vague. And for some reason the section heading is not appropriately capitalized.
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/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 — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The content is mostly good. There is some concern about the potential reproducibility of the work. While the tables show relevant information clearly, the same cannot be said about some of the figures. I cannot fully judge the content of the figures because there is a lot of information masked by the small print, which becomes too blurry when I try to enlarge it on the screen.
- 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 #2
- Please describe the contribution of the paper
The proposed Latent-CURE framework enhances breast ultrasound diagnosis by shifting away from standard end-to-end prediction toward a transparent, step-wise reasoning process. It aligns the model’s diagnostic trajectory with standard BI-RADS descriptors, sequentially analyzing the margin, calcification, and orientation before concluding. To combat dataset imbalance, the method couples an implicit chain-of-thought in the latent space with a Dual-Asymmetric Loss (Dual-ASL) mechanism. This innovative combination ensures that rare but critical malignant features are actively preserved to inform the final diagnosis.
- 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.
1) I like the overall narrative and structure of the paper; it is easy to follow and well-motivated. 2) While latent CoT has been explored before, combining it with a structured, multi-step feature-based reasoning pipeline is interesting and fits well with the clinical setting. 3) The Dual-ASL formulation is well-motivated for long-tailed clinical distributions and directly addresses the imbalance issue in a principled way.
- 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.
1) Can you expand upon the selection of feature trajectory i.e. margin → Calcification → orientation and then diagnosis? Are these permutations invariant? Also What is the contribution of each feature? Is there any ablation regarding that? 2) What is the level of annotation regarding the reasoning steps? 3) Many design choices are limited to breast cancer diagnosis. Can authors clarify if this framework is limited to breast cancer diagnosis? Have you conducted any experiment which proves the generalization capability? 4) It is not clear from the paper what exactly is the contribution of individual components—feature-based reasoning or dual asymmetric optimization. What are the numbers without Dual-ASL (is it row 2 from Table 3)? If so, why are they so low even compared to the Qwen baselines? Also from Table 2, does Qwen3-VL-Embed-COT actually have 0 specificity? 5) I like the overall flow of the paper, but there are instances where key details need to be much more clear—for instance, what exactly are these anchors and how they are selected? Or in Sec 3.3, which model is used for direct SFT? 6) How are anchor texts and queries constructed for each step? Please provide the exact phrases, tokenization approach, and whether alternative phrasings were tested.
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Although this paper has merit, my current assessment is a weak reject. This decision is based mainly on the limited novelty of the approach and several ambiguities in the writing. I am open to reconsidering after reading the authors’ rebuttal.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
I thank the authors for their rebuttal. While the method adds genuine value, I remain unconvinced that all three claimed contributions are independently proven. The latent reasoning conceptually parallels Process Reward Modeling (PRM), making Dual-ASL—which protects minority features—the true core contribution. Furthermore, it is unclear why standard CoT collapses so drastically without Dual-ASL, implying the reasoning module is entirely dependent on it. The paper would be significantly strengthened by adding ablations for the reasoning order and hyperparameters (w,m), alongside generalizable takeaways for other datasets. I lean towards a Weak Reject, but would not mind an Accept (would suggest the authors to integrate the rebuttal clarifications into the final manuscript. There is sufficient space to achieve this with minor restructuring).
Review #3
- Please describe the contribution of the paper
The paper proposes Latent-CURE, a breast ultrasound diagnosis framework that reformulates multimodal large model (MLM) inference as a deterministic metric learning problem in latent space rather than open-ended text generation. The three core contributions are: (1) an implicit Chain-of-Thought (CoT) module that forces sequential alignment with BI-RADS morphological descriptors (margin > calcification > orientation > diagnosis) before producing a final label; (2) a Dual-Asymmetric Loss (Dual-ASL) that applies feature-specific dynamic margins and sample weights to protect rare malignant anchors from gradient starvation under long-tailed class distributions; and (3) empirical validation on a 666-case IRB-approved multicenter BUS dataset, where the method achieves 93.78% accuracy and 87.07% MCC, outperforming both pure vision classifiers and general-purpose MLMs.
- 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.
1) Clinically grounded reasoning structure. Encoding the BI-RADS reasoning sequence directly into the latent trajectory is a meaningful design choice. Rather than treating clinical interpretability as a post-hoc explanation, the CoT pathway is architecturally enforced, which is a genuine departure from most end-to-end approaches. 2) Deterministic output via metric learning. Replacing probabilistic decoding with cosine nearest-neighbor retrieval over predefined semantic anchors is an elegant solution to the hallucination problem in medical MLMs. The theoretical guarantee of zero-hallucination outputs (within a bounded anchor set) is well motivated. 3) Dual-ASL is well-targeted. The combination of dynamic angular margin and asymmetric gradient weighting addresses a real and well-documented failure mode (gradient starvation on minority clinical features). The ablation in Table 3 compellingly demonstrates that the Dual-ASL is indispensable, removing it causes specificity to collapse from 97.16% to 4.26%.
- 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.
1) Small and single-modality dataset. With only 666 cases (approximately 533 training / 133 testing), the dataset is quite limited for a paper claiming state-of-the-art on a general clinical problem. Standard deviations are reported across seeds, but no cross-validation is performed. Generalizability claims are difficult to substantiate, and the potential for overfitting to this specific cohort is real. The paper does not report the benign/malignant split, which is critical context for evaluating MCC and specificity figures. 2) Dual-ASL hyperparameters appear hand-tuned without principled justification. Table 1 lists specific values (e.g., w=6.0, m=0.5 for malignant final diagnosis; w=4.0, m=0.4 for non-parallel orientation) but provides no ablation over these values and no description of how they were selected. It is unclear whether these were derived from held-out data, domain knowledge, or grid search, raising concerns about reproducibility and potential data leakage in hyperparameter selection. 3) The ‘implicit CoT’ framing is somewhat misleading. The latent trajectory is not truly a chain-of-thought in the generative sense, it is a fixed sequence of binary retrieval steps over predefined anchor pairs. This is closer to a cascaded multi-task metric learning setup. The distinction matters for claims of interpretability: the model outputs one of two pre-specified text labels per step, not a free-form rationale. Authors should clarify this distinction more precisely.
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/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 — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Latent-CURE addresses a clinically relevant problem with a technically interesting approach. The reformulation from generative decoding to latent-space metric retrieval is novel and well-motivated, and the Dual-ASL ablation provides convincing evidence of the loss function’s importance. However, several issues prevent a confident acceptance recommendation. The dataset is small (666 cases) with an unreported class split. The Dual-ASL hyperparameters appear hand-tuned without ablation or principled derivation. The comparison with frontier MLMs conflates zero-shot and fine-tuned conditions. The fixed CoT ordering lacks empirical justification.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors have addressed most of the concerns and clarifications I asked for - including the dataset split and training setup.
Author Feedback
We thank the chairs and reviewers for their review. We are encouraged that R1/R3 recognize Latent-CURE’s contribution and performance, and that R2 finds it well-motivated and technically interesting. Remaining concerns involve novelty, design clarity, scope, and presentation. We provide detailed responses below. We thank R1 for highlighting presentation issues, and will revise typos, formatting, figure readability, and discussion clarity in the final version.
Contribution/CoT framing (R2 & R3). Latent-CURE is not free-form generative CoT decoding rationales. Sec.2.1 reformulates diagnosis as deterministic cosine retrieval in normalized latent space. The contribution is coupling clinically ordered BI-RADS latent-anchor retrieval with feature-conditioned Dual-ASL: the former gives bounded intermediate evidence, the latter protects high-specificity anchors under imbalance.
Anchors/annotations/phrase robustness (R2). Supervision is descriptor-anchor level, not free-text rationale annotation. Intermediate steps use binary BI-RADS descriptors distilled after physician consultation (Sec.3.1); final labels come from surgical histopathology. Fixed anchor phrases are used in training/testing: “Not circumscribed margin”/”Circumscribed margin”, “No suspicious calcification”/”Suspicious microcalcifications”, “Parallel orientation”/”Non-parallel orientation”, and “Malignant”/”Benign”. In Eq.(3), q_t denotes step-specific retrieval context: implementation uses fixed user text “CoT”, while step identity comes from the active anchor pair and H_{t-1}, not freely varied query wording. Anchors/queries use Qwen3-VL-Embed-8B tokenizer/encoder. Phrase perturbation is valid; prompt-robustness studies test character-, word-, sentence-, and semantic-level variants, but this is not submitted evidence, so we do not claim prompt-invariance. Unlike open-ended prompting, Latent-CURE targets a controlled medical workflow where training/deployment share standardized BI-RADS anchors; phrase mismatch is constrained.
Feature trajectory/components (R2 & R3). The order margin->calcification->orientation->diagnosis is not permutation-invariant. Sec.2.2 defines it as a clinical progression from macro-structural cues to micro-signals and diagnosis. Since each step conditions on previous anchors through H_{t-1}, z_t=f_theta(X_v+q_t+H_{t-1}) changes if the order changes. We do not claim global optimality; the Discussion lists sequence changes and feature contribution as future questions. Table 3 separates modules: Direct SFT gives 91.79% Acc / 96.45% Spec / 83.26% MCC but is opaque; Standard CoT (“without Dual-ASL”) drops to 62.69% / 4.26% / -3.95%; full Latent-CURE gives 93.78% / 97.16% / 87.07%. Thus, Standard-CoT collapse is the failure mode Dual-ASL addresses: descriptor reasoning alone can leave high-specificity anchors under-optimized and biased toward majority-prior collapse. Table 2 reports Qwen3-VL-Embed-COT specificity as 0.00+/-0.00. Dataset split/metrics/scope (R2 & R3). The claim is confined to the submitted BUS setting: 666 IRB-approved multicenter cases, 8:2 split, and histopathology-derived final labels. The test subset has 87 malignant and 47 benign cases; accuracy alone can mislead, because an all-malignant predictor obtains 64.93% accuracy but 0 specificity, matching Qwen3-VL-Embed-COT in Table 2.Thus we emphasize specificity and MCC. We do not claim external validation or generalization to all cancers; new tasks require task-specific anchors/workflows.
Dual-ASL weights/margins (R3). Sec.2.3 states that w and m are conditioned on epidemiological prior and diagnostic specificity, and Table 1 discloses all values. These values were fixed a priori from physicians’ clinical prior on descriptor rarity/specificity; they were not derived from training-set descriptor prevalence, not selected by grid search, and not fitted on the test set. Thus Table 3 is module ablation rather than hyperparameter ablation; the controlled mechanism evidence is Table 3.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This paper presents Latent-CURE, a dual-asymmetric chain-of-thought framework for interpretable breast cancer diagnosis using latent space reasoning and BI-RADS-aligned clinical workflows. All three reviewers acknowledge the work’s meaningful methodological contributions, strong quantitative performance on a multicenter breast ultrasound dataset, and the clinical value of its structured, interpretable inference pipeline. Key shared weaknesses include problematic presentation quality, insufficient ablation studies, unclear design rationales, limited dataset scale, and ambiguous generalizability across broader cancer tasks. Reviewers hold mixed recommendations, with two weak accept votes and one weak reject decision heavily contingent on thorough author rebuttal to address technical and writing ambiguities. Overall, the work demonstrates solid technical innovation and clinical motivation but requires targeted revisions to resolve unclear details, improve readability, and validate component-specific contributions for final acceptance.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
There are still concerns on the effectiveness of each proposed innovation and potential reproducibility of the work.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Two reviewers support acceptance (R1=4, R3=4→Accept). R2 maintains Reject but states “would not mind an Accept” — a borderline case. The dual-asymmetric chain-of-thought approach for interpretable breast cancer diagnosis addresses a clinically important problem.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This paper introduces Latent-CURE, an interpretable diagnostic framework that reformulates open-ended generative decoding into a deterministic metric retrieval process within a protected continuous latent space to enforce a clinically grounded BI-RADS reasoning trajectory (Margin -> Calcification -> Orientation) for imbalanced breast ultrasound diagnosis. To protect rare but highly specific malignant features from gradient starvation against abundant benign priors, the authors couple this latent pathway with a novel Dual-Asymmetric Loss (Dual-ASL) optimization strategy that dynamically tunes angular margins and sample weights. While a negative recommendation persists due to concerns regarding domain specificity, structural dependencies on Dual-ASL to prevent standard chain-of-thought collapse, and blurry visual resolutions in some figures, these limitations are convincingly mitigated. Backed by an exceptionally solid empirical validation using an IRB-approved multicenter dataset of 666 pathologically verified cases where Latent-CURE achieves a state-of-the-art diagnostic accuracy of 93.78% and protects specificity from class-imbalance collapse.