Abstract

Narrow-Band Imaging (NBI) endoscopy plays a central role in laryngeal lesion screening, yet fine-grained risk stratification remains challenging due to strong visual overlap, vascular variability, and contextual bias. While single-stage detectors achieve reliable localization, they often struggle to structure lesion representations for robust classification, particularly for morphologically ambiguous and underrepresented categories as Leukoplakia. We propose a dual-branch real-time detection framework that integrates prototypical learning into a YOLO-based architecture to explicitly organize the latent space around clinically meaningful lesion prototypes. A multi-scale feature fusion module and an auxiliary prototype head are jointly optimized through a two-stage training strategy with bootstrap stabilization and exponential moving average prototype updates. Experiments conducted on a retrospective NBI cohort and an independent prospective dataset demonstrate that our approach significantly improves fine-grained stratification, achieving superior macro F1-score compared to state-of-the-art CNN- and Transformer-based detectors, while preserving competitive detection performance and real-time efficiency. The source code is publicly available at https://github.com/lorenzo-federici/Prototypical-YOLO.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5194_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/lorenzo-federici/Prototypical-YOLO

Link to the Dataset(s)

N/A

BibTex

@InProceedings{FedLor_Leuko_MICCAI2026,
        author = { Federici, Lorenzo AND Fiorentino, Maria Chiara AND Rosati, Riccardo AND Sampieri, Claudio AND Peretti, Giorgio AND Mora, Francesco AND Bellini, Elisa AND De Mattos, Leonardo S. AND Zingaretti, Primo AND Moccia, Sara},
        title = { { Leuko, I Am Your Prototype: A Prototypical Detection Framework for Fine-Grained Laryngeal Lesion Stratification } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16885},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The authors built a real-time framework based on YOLO11 nano with a “Prototypical Learning” twist. They used RoIAlign to map lesions into a feature space and created three learnable “anchors” (SCC, LEUKO, LR).

  • 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 is the first time integrating Prototypical Learning into a lightweight YOLO.

    2.Prototypical learning can be unstable. To fix this, the authors proposed a Bootstrap phase and EMA updates.

    3.The authors brought in a Prospective (PROS) cohort for external testing.

  • 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.YOLO lives by Mosaic and MixUp augmentation. It’s the only way to stop overfitting. However, the authors turned off Mosaic/MixUp after only 40 epochs. How did the authors actually handle the overfitting?

    2.YOLO needs to multiply objectness by class probability to run NMS. This work has two separate heads: one for YOLO, one for Proto. Which head’s class is used for the final box? If the Proto head only does classification, how is it synced with the YOLO head during NMS?

    3.YOLO’s whole point is being single-stage and dense. After integrating RoIAlign into the Proto Head, the algo becomes a two-stage move. Where do the boxes for RoIAlign come from during testing?

    4.There is zero data on Latency or FPS.

  • 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?

    clinical application

  • 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



Review #2

  • Please describe the contribution of the paper

    This paper proposes a dual-branch YOLO-based detection framework augmented with a prototype learning head to improve fine-grained classification of laryngeal lesions in NBI endoscopy.

    The method explicitly structures the embedding space using class prototypes and introduces a two-stage training strategy with bootstrap initialization and EMA prototype updates to improve discrimination, particularly for difficult classes such as leukoplakia.

  • 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 addresses a clinically important and challenging problem, namely fine-grained lesion stratification under strong visual ambiguity. The integration of prototype learning into a detection framework is well-motivated and practically meaningful. The proposed framework is carefully designed, with clear separation between detection and representation learning. The use of a bootstrap training phase is effective and well justified, improving stability of prototype learning. The method demonstrates strong improvements in classification performance, especially for minority and difficult classes (e.g., LEUKO). The evaluation includes both retrospective and prospective datasets, which strengthens clinical relevance and generalization. The model maintains real-time efficiency, which is important for deployment in endoscopic workflows.

  • 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 novelty is moderate, as prototype learning has been widely explored, and the contribution mainly lies in adapting it to YOLO. The method relies on carefully tuned training strategies (bootstrap, EMA, thresholds), which may limit robustness and reproducibility. Improvements are primarily in classification (F1), while detection gains are relatively small. The number of classes is limited (3 classes), which restricts demonstration of scalability. The dataset, although clinically collected, is relatively small and imbalanced, which may affect generalization claims. The method does not include comparison with other representation-learning strategies (e.g., contrastive learning).

  • 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 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.

    (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 paper presents a solid and practically relevant improvement for lesion stratification, with clear gains in challenging classes and strong experimental validation. While novelty is moderate, the work is well-executed and clinically meaningful, supporting a Weak Accept.

  • 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



Review #3

  • Please describe the contribution of the paper

    This paper proposes a dual-branch real-time detection framework for fine-grained laryngeal lesion stratification in Narrow-Band Imaging endoscopy. Built upon YOLO11n, the method adds an auxiliary prototypical learning branch to the standard detection head, using multi-scale feature fusion, RoIAlign, and a cosine-similarity-based prototype constraint to explicitly organize lesion representations around three clinically meaningful prototypes: SCC, LEUKO, and LR. The authors further design a two-stage training strategy consisting of a bootstrap stage with detection-only optimization followed by joint optimization of detection and prototype losses, with class prototypes updated via EMA. Experiments on a retrospective RETRO cohort and an independent prospective PROS cohort show that the approach improves macro F1 while preserving competitive detection performance, with the largest benefit observed for the difficult and underrepresented leukoplakia category.

  • 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 clinical problem is important and well targeted. The paper addresses risk stratification of laryngeal lesions in NBI endoscopy rather than lesion localization alone, which is more directly aligned with real clinical decision-making. The method is purposefully designed. Instead of simply appending another classifier, the authors use prototypical learning to explicitly structure the embedding space, with the goal of reducing confusion among morphologically similar classes and mitigating background/context bias. The gain on the difficult LEUKO class is notable. On RETRO, the Proposed (Proto) model reaches an LEUKO F1 of 0.42, clearly above YOLO11nano (0.26) and RT-DETR (0.15); on the prospective PROS cohort, it still achieves 0.40, outperforming competing methods. This is practically meaningful for a minority and clinically challenging class. The evaluation is relatively comprehensive. The paper includes comparisons with both CNN- and Transformer-based detectors, ablations on bootstrap epochs, backbone adaptation strategies, and the contribution of the prototype head, as well as an external prospective test set. The framework remains reasonably efficient. The authors state that the model still has only 7.45M parameters after adding the prototype branch, supporting its potential for real-time clinical deployment.

  • 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 end-to-end evaluation is not sufficiently clear. In Table 2, Proposed (Yolo) and Proposed (Proto) are reported in separate rows, and the Proto row does not include detection metrics. This makes it unclear whether the final classification results are based on GT boxes, predicted boxes, or some post-processing pipeline. For a framework claiming joint detection and stratification, this point should be explained more explicitly. The degree of “fine-grained” stratification is still limited. The current task uses only three classes, with all low-risk lesions grouped into a single LR category. This somewhat reduces the granularity of the task and makes the claim of fine-grained stratification slightly overstated. The external prospective dataset is relatively small. PROS includes only 327 images from 64 patients, with a distribution heavily skewed toward SCC. While having an independent prospective cohort is a strength, the scale still limits the robustness of the conclusions. Comparison against more direct imbalance-aware or discriminative baselines is limited. The paper mainly compares against different detector families, but it does not fully rule out whether the gain comes primarily from handling minority classes or enforcing a more discriminative embedding space, rather than from the specific prototype-head design itself. Comparisons to metric-learning, center-loss, class-balanced, or focal-style classification enhancements would strengthen the claim. The statistical testing section could be more rigorous. The authors use image-level bootstrap and describe the comparisons as independent pairwise tests, but the models are evaluated on the same test sets. The assumptions and pairing strategy 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 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?

    I recommend that the authors explicitly add an ethics approval, informed consent, or waiver statement in the manuscript to remove ambiguity regarding compliance for the patient data used.

  • 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?

    I assign this paper a weak accept. The main reason is that it presents a reasonable and fairly complete framework for a clinically relevant problem, namely fine-grained laryngeal lesion stratification in NBI endoscopy, and shows reasonably consistent classification gains on both retrospective and independent prospective cohorts, especially for the difficult minority LEUKO class. The paper also includes meaningful ablations supporting the role of the bootstrap strategy and the prototype head. At the same time, the weaknesses are also clear: the end-to-end evaluation protocol is not fully transparent, the external test set is limited in size, and the ethics/statistical reporting could be strengthened. Overall, I consider the paper slightly above the acceptance threshold, but not strong enough for a confident accept.

  • 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

N/A




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • 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 received positive feedback from all three reviewers. The integration of prototypical learning into YOLO is novel and practical, with classification improvements especially for challenging classes. The use of a prospective testing cohort further strengthens its clinical relevance. Nonetheless, reviewers raised concerns about the lack of clarity in overfitting mitigation, how detection and classification heads integrate during inference, and limited evaluation transparency. The scalability and novelty could also be improved by including more classes and broader comparisons with alternative methods. In addition, one reviewer suggested that authors explicitly add an ethics approval regarding compliance with the patient data used.



back to top