List of Papers Browse by Subject Areas Author List
Abstract
Cervical cancer remains a major global health challenge, with early diagnosis relying on accurate classification of cervical cell images. However, current computer-aided diagnosis systems often struggle with subtle morphological variations, overlapping cells, and imaging artifacts, particularly in ambiguous samples, due to their focus on single-cell views without contextual background analysis. Moreover, the lack of transparent reasoning in these models limits their interpretability. To address these issues, we propose CerviThink, an innovative framework that integrates reinforcement learning (RL) to support structured, context-aware diagnostic reasoning. CerviThink employs a large vision-language model, fine-tuned on a Retrieval-Augmented Generation constructed CerviCoT dataset. We construct a decoupled grounding-answering pipeline: the grounding phase uses focus, transform, and ignore operations to isolate key cellular features and mask suspected abnormal cells, while the answering phase optimizes classifications with the Dynamic Visual Hybrid Reward (DVHR). DVHR integrates the consistency reinforcement reward and background normality confirmation reward, enhancing interpretability through iterative refinement and uncertainty resolution. Extensive experiments demonstrate CerviThink’s superior performance over existing methods, underscoring its potential to advance computational pathology through context-aware, interpretable visual reasoning. The code is publicly available at https://github.com/feimanman/CerviThink.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5232_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/feimanman/CerviThink
Link to the Dataset(s)
ComparisonDetector dataset: https://github.com/kuku-sichuan/ComparisonDetector
HiCervix dataset: https://zenodo.org/records/11081816
BibTex
@InProceedings{FeiMan_CerviThink_MICCAI2026,
author = { Fei, Manman AND Jiang, Haotian AND Yi, Zhenyu AND Wang, Qian AND Zhang, Lichi},
title = { { CerviThink: A Reinforced Visual Reasoning Framework for Cervical Cancer Cell Classification } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16879},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents CerviThink, a reinforcement learning-based visual reasoning framework for five-class cervical cell classification. The paper also constructs a RAG-based Chain-of-Thought dataset, CerviCoT, to inject domain-specific diagnostic rationales derived from PubMed literature.
- 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.
This paper involves a substantial amount of work. The authors do not only propose a training framework, but also build a task-specific dataset, CerviCoT, to support the reasoning process in cervical cell classification.
Another strength is that the paper introduces several task-motivated modifications that are reasonable for this problem setting. The grounding-answering decomposition, the focus/transform/ignore operations, and the DVHR design are all aligned with the intuition of how difficult cytology cases may need to be examined more carefully.
- 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.The paper does not sufficiently establish that the RAG-generated rationales truly reflect pathologists’ diagnostic reasoning. CerviCoT is constructed by retrieving PubMed literature and generating step-by-step CoT rationales. This may be useful as weak supervision, and the ablation suggests it improves performance. However, the paper does not provide clinician verification, faithfulness assessment, or any expert agreement analysis showing that these rationales actually reconstruct how pathologists reason in practice. At present, the paper demonstrates utility of synthetic rationales, but not their fidelity to real diagnostic reasoning. 2.The methodological novelty appears moderate relative to the underlying Ground-R1/GRPO paradigm. The paper clearly adapts ideas from Ground-R1, including grounded visual reasoning and RL-based optimization, and builds on Qwen2.5-VL with a task-specific pipeline. In my view, the main novelty lies in the domain adaptation, the ignore operation in the cytology setting, the CerviCoT dataset, and the DVHR design. These are potentially useful contributions, but the core algorithmic backbone remains close to existing grounded RL reasoning frameworks. As a result, the novelty is more incremental/task-specialized than fundamentally methodological. 3.Strong vision-only baselines are missing. The empirical comparison is largely against LVLM-family baselines. For this task, especially once the input has already been cropped into 256×256 cell regions, strong pure-vision baselines such as ComparisonDetector, ResNet/ViT patch classifiers, or MIL/detection-first pipelines would be highly relevant and likely competitive. Their absence makes it difficult to judge whether the reported gains reflect a real advance over established cytology pipelines, or mainly an improvement within the LVLM comparison set. 4.There are reproducibility and reporting gaps. Important details are missing or under-specified, including reward weighting rationale, transform/inpainting parameterization, training stability measures, few-shot protocol details, and the exact role of bounding boxes. The paper also states that no explicit bounding box annotations are required, yet bounding boxes appear to be used for data construction and cropping. This distinction should be clarified more carefully.
- 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.
- 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?
This paper addresses an important problem and proposes a reasonably well-motivated framework with clear task-specific adaptations for cervical cytology. The reported results are promising. However, I remain unconvinced that the current empirical evidence is strong enough for acceptance at this stage. My main concerns are: (1) the paper does not validate that the RAG-generated rationales genuinely reflect pathologists’ reasoning; (2) the novelty, while real at the application/task-design level, appears limited relative to the broader Ground-R1-style grounded RL framework; (3) the central DVHR design is not sufficiently justified, particularly regarding equal reward weighting and possible reward bias; (4) the absence of strong vision-only cytology baselines significantly weakens the empirical claim of task-level progress.
- 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 updated experiments can address my concern. It is recommended that the updated experimental results are included in the camera ready version if this work can be accepted.
Review #2
- Please describe the contribution of the paper
This paper proposes CerviThink, a reinforcement-enhanced visual reasoning framework for cervical cancer cell classification. Built upon Qwen2.5-VL, the method seeks to emulate pathologists’ diagnostic reasoning process through reinforcement learning. The overall framework adopts a decoupled grounding-answering pipeline: in the grounding stage, the model identifies and processes key diagnostic regions through dynamic visual operations such as focus, transform, and ignore; in the answering stage, the model performs final classification optimization based on the Dynamic Visual Hybrid Reward (DVHR). DVHR consists of a format reward, an accuracy reward, a consistency reinforcement reward, and a background normality confirmation reward. In addition, the authors construct a RAG-based chain-of-thought dataset, CerviCoT, to inject medical knowledge into the model. Experimental results show that CerviThink outperforms multiple baseline methods on the DST, ComparisonDetector, and HiCervix datasets.
- 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.This paper addresses a clinically meaningful and challenging task. Cervical cell classification is strongly affected by subtle morphological differences, cell overlap, and imaging artifacts. The authors focus on the limitations of existing CAD systems and general LVLMs in complex scenarios, particularly their insufficient performance and limited interpretability. The problem setting itself therefore has clear practical value. 2.The overall methodological idea is relatively novel and well motivated. Rather than treating the task as a standard image classification problem, the paper attempts to mimic the diagnostic workflow of pathologists by decoupling “localizing critical regions” from “making classification decisions,” thereby building a reasoning framework that is better aligned with the clinical analysis process. This grounding-answering design fits the nature of the task well. The design of the dynamic visual operations is insightful. In particular, the ignore operation is not merely used to mask out regions, but instead masks suspicious abnormal cells and examines whether the remaining background is normal, thereby simulating a degree of differential diagnostic reasoning. This design makes the method more tailored to medical scenarios than generic visual reasoning frameworks. 3.The experimental section reports not only the main results but also includes a certain level of ablation analysis. According to the paper, CerviThink achieves an F1-score of 74.80 on the DST dataset, outperforming multiple general-purpose and pathology-related baselines. In the few-shot settings on ComparisonDetector and HiCervix, it also shows clear improvements over the SFT and RL variants. The ablations on CerviCoT and the reward components further support the effectiveness of both the data construction strategy and the full DVHR mechanism.
- 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.The overall framework is clearly inspired by existing grounded visual reasoning and RL-based LVLM methods, especially works closely related to Ground-R1 [1]. In the current manuscript, the main incremental contributions appear to lie primarily in the cervical cell-specific data construction, the ignore operation, and the DVHR design. However, the paper does not yet sufficiently clarify whether these modifications constitute substantive methodological innovations or are mainly task-specific engineering adaptations. 2.The important experimental details remain unclear, including the class distribution of the DST dataset, whether the train/test split is performed at the patient level, the exact sampling protocol for the few-shot setting, whether the experiments were repeated across multiple random seeds, and the key hyperparameters used in reinforcement learning training. The current implementation details are too brief to support reproducibility. 3.The main tables report only single-run Precision, Recall, and F1-score, without standard deviations, confidence intervals, or statistical significance analysis. Given that this is a medical image classification task and may involve class imbalance, reporting only weighted F1 is not sufficient. Providing per-class results, a confusion matrix, and more clinically relevant metrics would better support the paper’s conclusions. 4.The paper’s claims regarding “interpretability” and “human-like reasoning” are stronger than the evidence currently provided. Although the method introduces structured outputs such as
, , and , and designs a visual reasoning pipeline, there is still insufficient qualitative analysis showing whether the model has truly learned clinically meaningful reasoning processes. For instance, the paper does not show whether the localized regions align with the areas of interest used by pathology experts, nor does it include failure case analysis or expert evaluation to support the claim of improved interpretability. A case study would be more appropriate. 5.Some of the assumptions behind the reward design may be overly strong, especially the background normality confirmation reward, which assumes that after masking the suspected abnormal cells, the remaining region should be classified as normal. In real pathological images, however, multiple abnormal regions may coexist, or the surrounding background itself may contain diagnostically relevant abnormal cues. Therefore, this assumption may not always hold. The paper does not sufficiently discuss the applicability range or potential limitations of this design. [1] Cao, M., Zhao, H., Zhang, C., Chang, X., Reid, I., & Liang, X. (2025). Ground-r1: Incentivizing grounded visual reasoning via reinforcement learning. arXiv preprint arXiv:2505.20272.
- 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.
- 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?
This paper addresses a clinically meaningful and challenging task, and the problem itself has strong practical value. Rather than treating it as a standard image classification problem, the paper introduces a framework that combines grounded visual reasoning with reinforcement learning to emulate the diagnostic process of pathologists. Through the decoupled grounding-answering design, the ignore operation, and the DVHR reward mechanism, the method aims to better handle complex cervical cell morphology. The overall idea is relatively novel and aligns well with the analytical logic of medical diagnosis. Experimentally, the method shows clear improvements over the baselines on the DST, ComparisonDetector, and HiCervix datasets, and the ablation studies also provide some support for the effectiveness of both the CerviCoT data construction strategy and the complete reward design. Overall, the method is reasonably well designed, the paper is clearly organized, and the work appears potentially inspiring and impactful, which is why I lean toward acceptance. That said, the current work still has several aspects that would benefit from further strengthening. First, the boundary between this method and prior grounded visual reasoning / RL-based LVLM work is still not sufficiently clear. At present, the main incremental contributions seem to lie more in task-specific adaptation to the cervical cell domain, while the paper does not yet fully justify whether these designs constitute sufficiently strong methodological innovation. Second, the experimental details remain somewhat incomplete: key information needed for reproducibility, such as the data split protocol, few-shot sampling strategy, random-seed repetitions, and RL training details, is limited. The main results also lack standard deviations, significance analysis, and more fine-grained class-level metrics. In addition, the paper’s claims regarding interpretability and human-like reasoning are currently supported more by methodological intuition than by direct evidence, as the paper does not yet provide sufficiently strong qualitative case studies, failure analysis, or expert evaluation to substantiate these claims.
- 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.
I appreciate the authors’ effort in clarifying several important points. In particular, the response helps better position the paper with respect to Ground-R1-like prior work, and the additional explanation of the ignore operation, transform operation, and DVHR makes the intended methodological contribution clearer. The clarification that the train/test and few-shot splits were conducted at the patient level is also important, and the added hyperparameter details partially address my reproducibility concerns. Overall, the rebuttal reduces some of my uncertainty regarding the experimental setup and the paper’s motivation.
That said, the rebuttal does not fully resolve my main concerns. The core novelty still appears somewhat limited, as the main incremental contributions remain concentrated on task-specific adaptation to the cervical cytology setting rather than a clearly established methodological advance beyond prior grounded visual reasoning and RL-based LVLM frameworks. In addition, several important points are addressed mainly through promised revisions, such as adding standard deviations, per-class metrics, and more detailed implementation descriptions, rather than through evidence already contained in the submitted manuscript. The claims about interpretability and clinically meaningful reasoning are also still not fully supported by strong qualitative or expert-based evidence.
Overall, I find the paper meaningful, reasonably well executed, and potentially impactful for an important medical application. Although the current version still has some limitations in terms of novelty justification, completeness of experimental reporting, and evidence for interpretability, the overall contribution is positive enough that I maintain my weak accept recommendation.
Review #3
- Please describe the contribution of the paper
Authors propose CerviThink, an innovative framework that integrates reinforcement learning (RL) to emulate pathologists’ contextual diagnostic 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.
1.Cervi cancer diagnostics is crucial. 2.Authors propose a novel method based on reinforcement learning. 3.Experiments are sufficient and reasonable.
- 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.Figure 1 could be improved to show the data flow, input and output of the system. 2.According to template, table caption is put on top of the table.
- 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?
Overll quality of this paper is good. Figure 1 should be improved to increase the readability.
- 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.
N/A
Author Feedback
We thank the reviewers for their insightful feedback. We address key concerns below: 1.Methodological Novelty vs. Ground-R1 (R1, R2) We respectfully clarify that CerviThink transitions from the passive spatial grounding of Ground-R1 to active sequential reasoning. Ground-R1’s static extractive cropping is insufficient for cytology, where dense overlapping demands hypothesis-driven deduction. We advance this in three ways: (1) Instead of merely locating a cell, we structurally redefine the action space using an Ignore operation to actively intervene in the visual state. By masking the suspected lesion, it forces the model to decouple the foreground and verify the local background morphology, simulating a clinical “rule-out” process absent in Ground-R1.(2) To handle micro-textures, our Transform operation utilizes dynamic zooming and contrast adjustments. This actively resolves fine-grained features like chromatin details that remain inherently ambiguous at Ground-R1’s global scale. (3) Finally, our DVHR shifts the optimization target from standard spatial grounding to enforcing logical consistency across these visual interventions. This ensures all reasoning steps conceptually align, overcoming the limitations of conventional models in complex cytopathological scenes. 2.Clinical Validation & Interpretability (R1, R2) During the development of CerviCoT dataset, we collaborated with pathologists to review the generated reasoning paths. Their assessment confirmed that the model’s outputs exhibit clinical fidelity and align with the Bethesda system workflows. 3.Missing Vision-Only Baselines (R1) During our comprehensive baseline evaluations, we assessed vision-only models on the DST dataset. Specifically, ResNet-50 and ViT-B/16 achieved F1-scores of 68.12% and 71.85%, respectively, compared to CerviThink’s 74.80% on the DST Dataset. Unlike vision-only models that overfit to background artifacts and struggle with borderline cases, our sequential interventions effectively isolate and verify morphological features. 4.Assumption on Background Normality (R2) We agree with this insightful observation. In scenarios with multiple lesions, masking a single cell does not guarantee a completely normal background. Currently, this reward functions as a soft regularizer that encourages local context verification, thereby preventing the model from relying solely on the most salient lesion. We will explicitly discuss this assumption as a limitation in the revision and outline our future work on multi-instance masking strategies for complex, multi-lesion samples. 5.Reproducibility & Hyperparameters (R1, R2) To prevent data leakage, the 4:1 train/test and few-shot splits were performed strictly at the patient level. We will update all tables with standard deviations and per-class metrics. Regarding the role of bounding boxes, explicit annotations were used solely for initial data preprocessing. During RL training and inference, while the model dynamically generates spatial coordinates to guide visual operations, these coordinates are never supervised by ground-truth bounding boxes. Instead, spatial localization is driven entirely by the downstream reasoning rewards. For the hyperparameters, we set G1=4 and G2=2.KL divergence constraints were excluded, the clipping range was set to ε=0.2, and all DVHR components were weighted equally (1.0). Exact transform and inpainting parameterizations will also be fully detailed in the revision. 6.Figures & Formatting (R3) We will redesign Figure 1 to explicitly illustrate the tensor data flow, including inputs, policy network routing, and dynamic visual operations. All table captions will be moved to the top per conference guidelines.
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.
A RL-based reasoning framework is proposed to enhance cervical cancer cell classification. The framework adopts a two-stage pipeline, including a grounding stage to identify key diagnostic regions and an answering stage to perform classification. Reviewers recognize that the paper is clinically motivated and results are promising. However, they also raise issues that should be addressed in the rebuttal. Particularly, R1 points out several issues on evaluation, and R2 suggests to further clarify the novelty over existing works and provide more details and analysis.
- 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.
After the rebuttal, the reviewers are aligned in their recommendation to accept the paper. Although several minor issues remain, the overall contribution is valuable and technically sound. AC agrees with the reviewers’ assessment and recommends acceptance. The authors should incorporate the updated results and relevant clarifications into the camera-ready version.
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.
All three reviewers agree to accept. I read the rebuttal and reviewers’ comments after the rebuttal. I think that the rebuttal solves main concerns, and found not inappropriate comments from the reviewers.
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.
The paper presents a well-motivated visual reasoning framework with task-specific designs, including CerviCoT, the grounding-answering decomposition, the ignore operation, and DVHR.
Although the novelty is somewhat incremental relative to existing grounded RL/VLM reasoning frameworks, the rebuttal clarifies key experimental details, patient-level splits, the role of the task-specific components, and additional experiments. Given the clinical relevance, promising empirical results, and generally positive post-rebuttal reviewer opinions, I recommend acceptance, while asking the authors to clearly discuss the remaining limitations on rationale faithfulness, interpretability, reproducibility, and comparison with strong vision-only baselines.