List of Papers Browse by Subject Areas Author List
Abstract
Automated polyp segmentation in colonoscopy continues to
pose challenges due to substantial appearance variations and indistinct polyp boundaries. Although emerging foundation models (FMs) such as DINOv2, SAM, and OneFormer demonstrate remarkable generalization capabilities, their direct transfer to the polyp segmentation task and deployment in real-time clinical settings are difficult due to a lack of large-scale labeled data and high computational demands. In addition, adopting multiple FMs together raises concerns, even though they encode complementary semantic and structural information. While lightweight models, including U-Net, PraNet and U-Net++, are computationally efficient, they often struggle to generalize across datasets due to limited representational capacity. To address this gap, we propose Lite-PolypInductor (Lite-π), a novel foundation model induction framework that significantly enhances lightweight polyp segmentation baselines. Our proposed framework generates FM-specific prototype representations and aligns them semantically with the corresponding foundation model priors through reconstruction-based supervision. Subsequently, transformer-based fusion is introduced to highlight the polyp-relevant representations, particularly salient boundary information, while preserving complementary semantic cues. Extensive experiments across five polyp segmentation benchmark datasets demonstrate that Lite-π significantly improves lightweight baselines, achieving superior generalization performance with minimal computational overhead and thereby, offering a practical solution for generalized polyp segmentation. The code is available at https://github.com/lostinrepo/Lite-Pi.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5591_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/lostinrepo/Lite-Pi
Link to the Dataset(s)
N/A
BibTex
@InProceedings{AgnShi_Induce_MICCAI2026,
author = { Agnihotri, Shivanshu AND Majhi, Snehashis AND Nayak, Deepak Ranjan AND Mahapatra, Dwarikanath AND Jha, Debesh},
title = { { Induce to Empower: Improving Lightweight Baselines via Foundation Model Induction for Generalized Polyp Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16881},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes Lite-PolypInductor (Lite-π), a novel foundation model induction framework that can enhance lightweight polyp segmentation baselines. The research content is in line with current research trends.
- 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 authors did not directly use the basic models such as SAM and DINOv2 as frozen backbones, nor did they simply distill the features of large models to small models. Instead, the lightweight model generates synthetic prototypes that correspond to foundation priors internally, and gradually absorbs various prior information through reconstruction alignment. Furthermore, these multi-source priors are not simply concatenated, but first transformed into lightweight models that can generate and control internal representations, and then selectively fused through Transformer and Mamba.
- 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 experimental results are not sufficient enough, and there is a lack of evidence to prove why UFMI is superior to conventional distillation. The ablation experiment is incomplete, only the ablation experiment results of U-Net, U-Net++, and PraNet have been replaced, and the ablation experiment results of key modules are missing. This study emphasizes the suitability for clinical real-time deployment, but does not report metrics such as inference speed (FPS), latency, GPU memory usage, etc. Only parameters/FLOPs are not sufficient to prove real-time performance. In addition, lightweight and easy to deploy models such as U-Net, U-Net++, PraNet, etc. are used in this paper. Whether they can truly improve generalization ability without significantly increasing deployment burden needs to be verified with experimental data.
- 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 experimental results in this paper are not sufficient enough. The ablation experiment is incomplete, only the ablation experiment results of U-Net, U-Net++, and PraNet have been replaced, and the ablation experiment results of key modules are missing.
- 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
This paper proposes Lite-PolypInductor (Lite-π), a foundation-model induction framework for generalized polyp segmentation. Instead of directly deploying large foundation models at inference time, the method uses three frozen foundation models—SAM, DINOv2, and OneFormer—to guide lightweight baselines through a Unified Foundation Model Inductor (UFMI). The UFMI generates FM-specific synthetic prototype embeddings from lightweight encoder features, aligns them with FM-derived priors via a reconstruction objective, and then fuses them using transformer- and Mamba-based refinement. The overall goal is to improve the generalization ability of lightweight segmentation models with limited extra cost.
- 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.Practical and well-motivated problem setting. The paper addresses an important gap between strong but heavy foundation-model-based methods and efficient lightweight baselines for polyp segmentation. This is a practically meaningful direction for real clinical deployment. 2.A technically coherent framework. The proposed UFMI is a reasonable and interesting design: it does not simply distill features from multiple foundation models, but introduces FM-specific prototype generation and semantic alignment, followed by transformer/Mamba-based fusion. This gives the method a clear architectural identity. 3.Strong and broad empirical evaluation. The method is evaluated on five benchmark datasets, including seen and unseen settings, and consistently improves three lightweight baselines (U-Net, U-Net++, PraNet). This makes the paper stronger than many works that only validate on a narrow setup. 4.Good positioning against relevant prior work. The paper directly compares against both conventional SoTA methods and the closest prior lightweight FM-transfer framework, Polyp-DiFoM, and shows consistent improvements. This comparison is important and strengthens the empirical contribution. 5.Efficiency-oriented design is appealing. Compared with heavy FM-based approaches such as SAM-Mamba and SAM-MaGuP, Lite-π achieves competitive performance with much smaller parameter/FLOP requirements on lightweight backbones, which is a meaningful contribution from an application perspective.
- 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 novelty is moderate rather than highly groundbreaking. The paper is closest in spirit to recent FM-transfer/distillation-style approaches such as Polyp-DiFoM, and the contribution mainly lies in the induction/prototype-alignment formulation and the specific multi-FM integration strategy, rather than in a fundamentally new paradigm. 2.The ablation study could be deeper. The current ablation mainly studies the effect of progressively adding multiple FMs. This is useful, but it would be even stronger to isolate the contribution of the core UFMI components more explicitly, e.g., prototype generation, reconstruction alignment, transformer fusion, and Mamba layer. 3.Efficiency claims would be stronger with runtime evidence. Although parameters and FLOPs are reported, the paper would benefit from actual inference speed, latency, or memory usage measurements if the authors want to emphasize real-time clinical applicability.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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.
(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 recommend Accept. The paper addresses a relevant and practical problem, proposes a technically sound multi-FM induction framework, and supports its claims with relatively broad experiments across multiple datasets, seen/unseen settings, and multiple lightweight baselines. In my view, the strongest aspect of the work is not extreme methodological novelty, but rather the combination of solid design, clear practical motivation, and convincing empirical validation. The paper is well positioned for the generalized polyp segmentation setting and demonstrates a favorable balance between accuracy and efficiency.
- Reviewer confidence
Very confident (4)
- [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.
In the rebuttal, the author has provided further explanations for the questions I raised, including methodological innovation, deficiencies in ablation experiments and detailed data on computational efficiency. Therefore, I maintain my decision to accept this paper.
Review #3
- Please describe the contribution of the paper
This manuscript proposes Lite-π, a generalized polyp segmentation framework. The method improves lightweight baselines by introducing knowledge from multiple foundation models, namely SAM, DINOv2, and OneFormer. Specifically, it generates FM-specific prototype embeddings from the latent representations of a lightweight segmentation network and aligns them with FM-derived priors through a reconstruction-based objective. The aligned features are then fused using Transformer layers and Mamba Vision layers, and finally decoded by a lightweight backbone. The method is evaluated using three lightweight baselines on five benchmark polyp datasets, and demonstrates improved performance with relatively small computational overhead.This manuscript proposes Lite-π, a generalized polyp segmentation framework. The method improves lightweight baselines by introducing knowledge from multiple foundation models, namely SAM, DINOv2, and OneFormer. Specifically, it generates FM-specific prototype embeddings from the latent representations of a lightweight segmentation network and aligns them with FM-derived priors through a reconstruction-based objective. The aligned features are then fused using Transformer layers and Mamba Vision layers, and finally decoded by a lightweight backbone. The method is evaluated using three lightweight baselines on five benchmark polyp datasets, and demonstrates improved performance with relatively small computational overhead.
- 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.The authors conducted extensive experiments to evaluate the proposed method, which is commendable. The results show that Lite-π consistently improves the baselines across many major metrics. 2.The motivation of enhancing lightweight polyp segmentation models with richer priors from foundation models is reasonable, especially since it avoids the substantial inference cost of directly deploying FM-based segmentation systems. This is an important direction for clinically relevant and resource-efficient deployment. 3.The manuscript presents a clear architectural design. The overall pipeline, including the baseline-specific encoder, the unified foundation model inducer, and the FM-induced decoder, is clearly illustrated.
- 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.Although the motivation of enriching lightweight models with stronger priors from foundation models is reasonable, the overall contribution to polyp segmentation appears limited. In comparison with recent prior work, especially Polyp-DiFoM [1], the methodological novelty is limited. The overall problem formulation in this manuscript is highly similar to that of Polyp-DiFoM, as both aim to transfer knowledge from multiple foundation models to lightweight baselines for generalized polyp segmentation. This paper replaces distillation with prototype induction and reconstruction-based alignment, but the high-level paradigm remains essentially unchanged. The manuscript does not sufficiently clarify why this shift constitutes a substantial methodological advance, and it appears more like an incremental variation of Polyp-DiFoM. 2.Core ablation studies are lacking. The ablation experiments mainly investigate the number of merged foundation models, which is indeed worth exploring, but they do not analyze the contributions of the reconstruction alignment loss, the Transformer or MVL fusion, or the progressive weighting design. Such isolated ablations would help better identify the true sources of the performance gains. 3.The paper does not provide sufficient implementation details, especially regarding the loss weight settings, and it also does not provide an anonymous code link. 4.The fairness of the comparisons remains unclear. It is not fully specified whether all compared methods were evaluated under the same preprocessing pipeline, input resolution, training budget, and hyperparameter settings. 5.The experimental section lacks statistical reporting, such as variance/confidence intervals and significance testing. [1] Agnihotri S, et al. “From SAM to DINOv2: Towards Distilling Foundation Models to Lightweight Baselines for Generalized Polyp Segmentation.” WACV. 2026.
- 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.
- 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 the paper proposes enhancing lightweight polyp segmentation models with richer priors from foundation models and shows some empirical gains, its overall contribution to the polyp segmentation field appears limited. Compared with the most closely related recent work, the novelty of this study is limited. The manuscript adopts prototype induction and reconstruction-based alignment for prior enhancement, it does not provide sufficient conceptual or empirical analysis to establish this as a strong methodological advance. In addition, the ablation study is rather limited relative to the central claims of the paper, and the argument for real-time deployment is not adequately validated by direct runtime evidence.
- 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 sincerely thank the reviewers for the constructive feedback and encouraging comments. We are pleased that the primary objectives of Lite-π were recognized. We address the main concerns below.
Novelty- Induction vs. Polyp-DiFoM / Distillation (R2, R3): We thank the reviewer for highlighting Polyp-DiFoM is the closest prior work. While both methods leverage foundation model (FM) priors for lightweight polyp segmentation, Polyp-DiFoM follows a distillation paradigm: it extracts FM features, decomposes them into low-/high-frequency components using FFT, and aligns student latent features with these teacher representations through MSE losses. In contrast, Lite-π follows an induction paradigm. Instead of forcing the lightweight model to imitate external FM features, Lite-π reconstructs polyp-relevant FM’s prototype features from the model’s own latent space and aligns them with frozen FM priors through reconstruction supervision. Key Differences: Since the FM prototypes are generated from task-trained lightweight representations, Lite-π learns to retain task-relevant semantic and boundary priors while suppressing irrelevant, redundant, or noisy FM responses. This is ensured by jointly optimizing the segmentation loss and feature reconstruction loss, where the reconstructed FM features directly influence the segmentation representation. Therefore, only FM-induced features that are useful for polyp segmentation are reinforced. This functionality is naturally not available in conventional distillation, where the student may inherit irrelevant, redundant, or noisy teacher responses, potentially degrading segmentation.
Ablation Depth (R1, R2, R3): The current ablation study was designed to analyze the inductive effect of progressively incorporating multiple foundation models, since the proposed UFMI framework is designed as a unified induction module. The Transformer/MambaVision fusion, reconstruction alignment, and progressive weighting are tightly coupled within UFMI and jointly operate to perform semantic prototype induction and fusion. Therefore, the current ablation already validates the main mechanism: scalable multi-FM induction improves generalization without a significant computational burden. Nevertheless, we agree that additional component-level analysis could further improve clarity, and we will expand the ablation discussion in the revised manuscript to better explain the role and interaction of these tightly integrated components.
Runtime, Latency, and Deployment Cost (R1, R2): Due to page constraints, we only reported parameters and FLOPs. Lite-π already shows modest Parameters-FLOPs overhead: U-Net increases from 16.7M-73.9G to 19.6M-74.3G, U-Net++ from 9.1M-65.9G to 10.1M-66.6G, and PraNet from 30.4M-13.1G to 34.9M-13.9G. Compared with SAM-Mamba/SAM-MaGuP (>100M parameters and >400G FLOPs) [Tables 1–2 ], Lite-π remains substantially lighter while achieving competitive accuracy. These results support the low deployment burden of Lite-π; we will include additional FPS, latency, and GPU memory reporting where space permits.
Generalization with Low Deployment Burden (R1): Compared with the original baselines, Lite-π improves average mDice by +11.07% across all 15 dataset-backbone cases and by +13.32% on unseen datasets, demonstrating cross-dataset generalization with lightweight inference.
Fair Comparison, Implementation Details and Reproducibility (R3): We will explicitly state that all reproduced methods use the same dataset split, input size (352×352), augmentation, training budget, and hyperparameters. Further, we will provide source code upon acceptance.
Statistical Reporting (R3): Following standard practice in polyp segmentation, our manuscript provides a comprehensive evaluation across five benchmark datasets and six widely used metrics. The improvements shown are robust, reproducible, and not dataset-specific. We will clarify this strength in the revised manuscript.
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.
Please address the concerns about Incomplete ablation studies, missing runtime, limited methodological novelty.
- 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 accept and one reject.
i read through and would accept subject to minor
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.
The paper proposes Lite-π, an FM-induction framework for improving lightweight polyp segmentation baselines using priors from SAM, DINOv2, and OneFormer. The reviewers agree that the problem is clinically relevant and that the empirical evaluation is broad, covering five datasets, seen/unseen settings, and three lightweight baselines. The submitted results show consistent gains over the baselines and favorable comparison to Polyp-DiFoM and heavier FM-based methods with relatively small parameter/FLOP overhead. Concerns remain: the novelty over Polyp-DiFoM is moderate, the ablation study does not isolate all UFMI components, and runtime/memory evidence for real-time deployment is missing in the submitted paper. The rebuttal reasonably clarifies the conceptual distinction between induction and distillation and comparison protocols, though promises of additional analyses cannot be considered. Overall, despite limitations, the work is technically coherent, practically useful, and sufficiently validated for acceptance.
Meta-review #3
- 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.
Initially mixed reviews favoring an accept (5-4-3). Unfortunately, R3 (weak reject) did not participate in the post-rebuttal phase, as it would be helpful to see whether her/his concerns were addressed in the rebuttal. Also, the motivation of R1 seems a bit off considering to propose a weak accept.
Personally, I’m very on the fence about this paper, as the method is only evaluated on polyp segmentation, which is rather niche and knows a significant inter-observer variability (moreover, bounderies of polyps are often not so well defined, especially in endoscopy images). Therefore, the empirical results that support the claims of improved preformance hinge too much on a narrow and ambigious baseline application.
On the other hand, the technical contributions seems meaningful and interesting, while novelty may be limited, as indicated by R3. Given that R1 might have entered the wrong score, based on her/his comments, the valid concerns of R3 and my own impression of the paper, I feel that this paper not (yet) meets the bar for MICCAI and I suggest the authors expand their experiments with different (more meaningful and clear) segmentation tasks.