List of Papers Browse by Subject Areas Author List
Abstract
Federated learning (FL) is severely hindered by statistical heterogeneity due to variations in scanners, acquisition protocols, and patient populations. Such non-IID data induces client drift during local optimization, leading to unstable convergence and suboptimal global models when parameter-based aggregation is applied. We propose a prototype-based, influence-aware federated learning framework (FedProIn) that uses multiple learnable class prototypes to capture shared semantic structures across heterogeneous clients. We introduce feature divergence loss and prototype contrastive loss to mitigate client drift by decomposing it into feature drift and prototype drift. In addition, we propose a normalized influence aggregation strategy that adaptively weights client prototypes according to their contribution to the global representation, reducing the impact of biased or low-quality updates. Experimental results on two publicly available medical datasets, HAM10000 and Matek-19, demonstrate that FedProIn achieves accuracies of (83.5% IID, 81.1% non-IID) on HAM10000 and (96.2% IID, 95.8% non-IID) on Matek-19, respectively, outperforming existing baselines in both conditions. Our code is available at https://github.com/harsh-kmr/FedProIn.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1947_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/harsh-kmr/FedProIn
Link to the Dataset(s)
HAM10000: https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/DBW86T
MATEK19: https://www.cancerimagingarchive.net/collection/aml-cytomorphology_lmu/
BibTex
@InProceedings{KumHar_FedProIn_MICCAI2026,
author = { Kumar, Harsh AND Garg, Tarun Kumar AND Sundaresan, Vaanathi},
title = { { FedProIn: Mitigating Client Drift for Learnable Prototypes in Federated Medical Imaging } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a federal learning framework called FedProIn to solve the problems of long tail distribution and statistical heterogeneity in medical image data. This method decomposes client drift into ‘feature drift’ and ‘prototype drift’, and introduces feature divergence loss (FDL) and prototype contrast loss (PCL) to constrain them respectively. In addition, a normalized influence aggregation (NIA) mechanism is proposed to dynamically allocate aggregation weights according to the local activation frequency of the prototype. The authors verified the performance improvement of the framework in the Non-IID environment on the HAM10000 and Matek-19 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.
Strengths: (1) The motivation of the method is clear, and the illustrations are simple and easy to understand. (2) The method is effective in multiple datasets. (3) The authors captured the double drift problem in medical FL, and abstracted the client drift into feature space alignment and semantic center drift. The logical architecture of the method is clear.
- 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.
Weaknesses: (1) The authors claim that the method alleviates the feature drift, but does not provide any T-SNE or UMAP visual charts to prove that the feature distribution of different clients does achieve spatial alignment after the introduction of FDL. Limited ablation experiments alone cannot confirm the effectiveness of its alignment. (2) In the NIA part of the method, the prototype with very low activation times will still be given very low weight in the global aggregation. This is a logical conflict with the purpose of protecting rare features mentioned by the authors. (3) The authors did not count and show whether the prototypes of rare classes really gained the high weight as expected under the long tail data, or were still suppressed by the mainstream classes. (4) If the local rare sample of a client contains serious noise, NIA may give the sample a very high weight because of its ‘unique activation’ on a specific prototype, resulting in noise pollution of the global prototype. There are no test results for noise samples or low-quality updates. (5) The method uses M learnable prototypes to capture the heterogeneity of medical images. The paper does not show what semantic information these M prototypes represent. If the information captured by these prototypes is biological information or noise information, how will the generalization of the method be affected? (6) The ablation experiment showed that the removal of FDL decreased more than the removal of PCL, and the feature drift was the main failure mode. This conclusion was only statistically correlated, not causal.
- 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.
(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?
My recommendation is mainly based on experimental rigor and technical contribution effectiveness.
- 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.
I recommend accepting this paper. The authors have provided a convincing rebuttal that addresses the primary concerns raised during the initial review. They have clarified the normalization mechanism of NIA, demonstrating how it preserves rare class prototypes from being diluted without corrupting the global representation under noisy distributions. The authors have also provided an analysis of hyperparameter sensitivity, proving that the theoretical system overhead of FedProIn is negligible compared to FedAvg. I think this paper will be a valuable contribution to the MICCAI community.
Review #2
- Please describe the contribution of the paper
The paper proposes FedProIn, a federated learning (FL) framework that uses multi-prototype representation learning and influence-aware aggregation to address data heterogeneity across clients. Unlike prior methods that rely on single class means, FedProIn introduces multiple learnable class prototypes per class to better capture diverse semantic structures and enhance class separability. To mitigate client drift, the method explicitly decomposes it into feature drift and prototype drift, addressed through a feature divergence loss and a prototype contrastive loss, respectively. Additionally, the paper proposes a Normalized Influence Aggregation (NIA) mechanism, which adaptively weights client contributions by estimating the relevance of their prototypes, leading to more robust global aggregation. Experiments on two public medical imaging datasets demonstrate that FedProIn consistently outperforms existing FL baselines under heterogeneous settings.
- 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 problem statement is well motivated in the context of heterogeneous federated learning, and the proposed approach is conceptually grounded. The decomposition of client drift into feature drift and prototype drift is well-justified and reflects a clearer characterization of heterogeneity. Moreover, the influence-aware aggregation mechanism, which weights clients based on prototype utilization, is a reasonable extension for the adaptive aggregation approach.
The method shows strong empirical performance, outperforming multiple baselines, including prior prototype-based approaches, especially under high heterogeneity. The inclusion of statistical significance testing in ablations further strengthens the credibility of the various integrated components.
Furthermore, the paper provides a component-wise ablation of the multi-term loss (Eq. 4), offering clear insight into the contribution of each component to the overall performance.
- 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.While the proposed aggregation weights prototypes based on their utilization, the paper does not sufficiently demonstrate that this reliably captures true informativeness rather than usage frequency alone. 2.No information is provided on the mode architectures used for the experiments. Additionally, evaluating the method across multiple architectures would strengthen the generality of the proposed approach. 3.Although statistical significance testing is provided for ablation studies, it is absent for the main experimental results (Table 1). Reporting the significance testing for the main results would strengthen the empirical claims. 4.The value of M (number of prototypes per class) used in the main results (Table 1) is not explicitly stated. While optimal values are discussed in the ablation section, they should be clearly reported alongside the main results for clarity. 5.The paper claims that HAM10000 favors M = 2 prototypes per class. However, Figure 2 appears to indicate better performance at M = 1, leading to confusion that should be clarified. 6.The evaluation is limited to a fixed heterogeneity setting. As a result, it is unclear how the method behaves across varying degrees of heterogeneity (e.g., different Dirichlet parameters or number of clients).
- 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
I found the last paragraph of the “Server-side Aggregation” subsection somewhat unclear. In particular, it would be helpful if the authors could further elaborate on how the proposed Sparse Prototype Activation formulation (Eq. 7) simultaneously addresses both Limitation 1 and Limitation 2.A more detailed explanation or intuition would improve clarity.
- 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 addresses an important problem in federated learning under heterogeneous data and proposes a well-structured framework combining multi-prototype learning with influence-aware aggregation. The empirical results are strong and consistent, particularly in highly heterogeneous settings, where the method demonstrates clear improvements over relevant baselines.
However, the paper has several weaknesses in terms of clarity, experimental completeness, and validation that should be addressed. My inclination toward a weak accept is primarily driven by the strength of the empirical performance. Addressing the identified concerns would further strengthen the paper in terms of clarity, robustness, and overall impact.
- 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.
Most of my concerns have been addressed in the rebuttal. The only remaining issue relates to the discussion of statistical significance testing. While I understand the limitations imposed by the small number of runs, the response somewhat overstates the infeasibility of statistical testing, as alternative approaches such as paired t-tests or Wilcoxon signed-rank tests are commonly used in low-sample settings despite their inherent limitations. However, I do not consider this issue sufficient to outweigh the overall strengths of the paper. So I continue to accept the paper.
Review #3
- Please describe the contribution of the paper
In this paper, the authors proposed FedProIn, an influence-aware multi-prototype federated learning framework designed to tackle statistical heterogeneity and client drift in long-tailed medical imaging data. Specifically, it integrates learnable class prototypes and a drift decomposition strategy locally, alongside a normalized influence aggregation mechanism on the server side. Firstly, the local training process treats prototypes as optimizable parameters and explicitly decomposes client drift into feature drift and prototype drift, mitigating them through Feature Divergence Loss (FDL) and Prototype Contrastive Loss (PCL). Then, the server-side aggregation strategy utilizes a Normalized Influence Aggregation (NIA) mechanism to adaptively weight client prototypes based on their actual utilization during local training, which reduces the impact of noisy updates and preserves rare disease representations. Experimental results demonstrate that the proposed method achieves robust performance and outperforms existing federated learning baselines under both IID and non-IID conditions on public medical 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.
a)The paper explicitly decomposes client drift into “feature drift” and “prototype drift” via Feature Divergence Loss (FDL) and Prototype Contrastive Loss (PCL), offering a theoretically sound solution to statistical heterogeneity. b)The Normalized Influence Aggregation mechanism adaptively weights prototypes based on actual local utilization. This method effectively preserves rare disease features in long-tailed data and prevents them from being overwhelmed by common cases. c)The authors conduct comprehensive experiments on two medical datasets (HAM10000 and Matek-19). The results demonstrate consistent superiority over state-of-the-art baselines in both IID and non-IID settings, and the ablation study clearly confirms the necessity of all proposed components.
- 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.
a) The title emphasizes handling “long-tailed” data, but the framework relies on existing external techniques (class-balanced sampling and adaptive augmentation) to address class imbalance. The proposed modules (FDL, PCL, NIA) primarily solve client drift, making the “long-tailed” claim in the title misleading. b) The implementation details state that the loss weighting coefficients are “determined empirically.” The authors should clarify the rationale behind these specific values and discuss the potential impact if these values fluctuate in real-world deployments. c) FedProIn requires clients to maintain an extra frozen global encoder and transmit multiple prototype matrices. The paper completely lacks an analysis of the increased memory footprint, computational cost, and communication bandwidth compared to standard methods like FedAvg. d) Several equations throughout the paper (e.g., Equation 2) lack trailing punctuation marks. Please ensure all equations are properly punctuated as parts of their respective sentences.
- 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.
(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 recommend a “Weak Accept”. The proposed FedProIn framework is theoretically sound and empirically effective for handling client drift. However, to secure a full “Accept”, the authors should use the rebuttal phase to clarify the rationale behind their hyperparameter choices, discuss the model’s specific effectiveness on rare classes, and address the theoretical system overhead introduced by the multi-prototype design.
- 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.
The rebuttal confirmed the “long-tailed” claim was heavily overblown, and their rare-class evaluation completely lacks comparison with strong external baselines, so I am lowering my score to a Weak Reject.
Author Feedback
We thank the reviewers for their valuable feedback. NIA, Rare Prototypes & noisy distributions [R1:Q2-4, M-R(2)]: NIA is computed per prototype index across clients, not all class prototypes. In Eq.(7), each p_{global}^{k,c,t} is normalized only by \sum_{i \in \mathcal{S}_t} \mathcal{I}_i^{k,c}, so only clients activating prototype {k, c} contribute to its update. A rare-class prototype activated by one or two clients is aggregated from only those updates, not diluted by higher-frequency prototypes. Regarding noise, NIA may upweight a uniquely activated prototype, but this effect is confined to it and does not corrupt the full class representation. Clarified in the paper.
Semantic Meaning of Multiple Prototypes and Generalization [R1:Q5]: Learnable prototypes (LPs) are optimized as means of Gaussian mixture components per class in the latent feature space [2]. The task loss enforces intra-class compactness and inter-class separability [2], encouraging each prototype to converge toward a distinct clinical mode (e.g. pathology presentations), rather than noise or acquisition artifacts (which lack consistent class-discriminative structure), providing stable prototype positions for better generalisability.
UMAP [R1:Q1, M-R(1)]: Added UMAP plots for SOLO vs. FedProIn (non-IID) as Fig. 2, showing better overlapping feature distribution with FedProIn.
FDL association [R1:Q6]: Rephrased “confirming that …” to “indicate a stronger association …” acknowledging our ablation establishes statistical correlation rather than causality.
Prototype Informativeness [R2:Q1]: A prototype gets gradients only when chosen as nearest neighbor, refining frequently used ones for discriminability. If influence scores showed only frequency, NIA would boost generic prototypes, lowering performance. However, ablation (Fig. 3) shows NIA consistently outperforms standard aggregation, indicating informativeness.
Prototype Count [R2:Q4-5, M-R(3)]: Table 1 shows the best W-F1(79.3%) for both loss functions (MCEL and MCL), with M=2 under MCEL. Fig. 3 isolates MCL for ablation, where M=1 is optimal (W-F1: 77.0% vs. 75.4%). The apparent contradiction reflects with loss- and dataset-dependent optimal M, which we now state explicitly.
Model Architecture & Heterogeneity Settings [R2:Q2,Q6]: ResNet-18 is used, as stated in Sec. 3.Evaluating other architectures, Dirichlet parameters, and client counts remains future work.
Statistical testing [R2:Q3, M-R(3)]: We agree statistical testing strengthens empirical claims. Table 1 uses <5 runs, insufficient for reliable testing; Mann-Whitney U was feasible for the ablation due to 15 samples per variant (5 settings × 3 runs). However, Table 1’s consistent margins across both datasets and settings (std ≤0.5%) support reliability.
“Long-tailed” Title [R3:Q1, M-R(4)]: Given the test dataset distribution characteristics, we used “Long-tailed” in the title. Also, the proposed modules are not orthogonal to the long-tailed problem: by preserving minority class prototype representations during aggregation, they complement class-balanced sampling (suggested by tail-class F1 gain: Full: 65.3% vs. Ablated: 56.0%). However, since FDL, PCL, and NIA primarily target client drift, we will revise the title to ‘FedProIn: Mitigating Client Drift for Learnable Prototypes in Federated Medical Imaging’.
Hyperparameter Sensitivity [R3:Q2]: We tuned hyperparameters via sequential search on a 1,000-sample subset, sweeping one parameter at a time (added in the paper). For real-world robustness, we ran a log-scale sensitivity analysis on λ_FDL and λ_PCL, varying each by ±1 order of magnitude from their optimal values (HAM10000, Dir α=0.5, MCL) (Base:77.0%, 10× up:74.9%, 10× down:74.3%).
Resource Overhead [R3:Q3, M-R(5)]: FedProIn adds M×C×512 parameters (~0.5MB for M=2, C=7) and a tiny M×C influence matrix per round, negligible vs. FedAvg. The frozen global encoder adds no communication cost. Added in the paper.
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.
To move the paper toward acceptance, the rebuttal must focus on the following key points: (1) Provide evidence (e.g., t-SNE/UMAP) that feature divergence loss actually aligns feature distributions across clients. (2) Clarify the apparent contradiction in NIA: how it protects rare prototypes if low-usage prototypes still receive low weights—show real activation statistics for minority classes. (3) Address missing experimental details: number of prototypes M used in main results, model architectures, and statistical significance for Table 1.(4) Explain the “long-tailed” claim: how FDL/PCL/NIA specifically benefit rare classes beyond standard class-balancing techniques. (5) Provide a rough estimate of extra memory, computation, and communication costs compared to FedAvg.
- 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 authors’ rebuttal convincingly addresses the major concerns raised by the reviewers, including the working mechanism of NIA for rare prototypes, the semantic interpretability of multiple prototypes, and the resource overhead analysis. The paper makes a solid methodological contribution to federated learning for medical imaging, and the remaining concerns do not undermine its overall validity.
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 addresses an important problem in federated medical imaging, namely client drift under statistical heterogeneity, and proposes a coherent multi-prototype framework to handle both feature-level and prototype-level drift. The decomposition into FDL and PCL is well motivated, and the Normalized Influence Aggregation mechanism provides a reasonable way to aggregate prototypes according to their actual client-side usage. The empirical results on HAM10000 and Matek-19 are consistently strong, and the ablation studies support the contribution of the main components. The reviewers raised valid concerns about the evidence for feature alignment, the behavior of NIA on rare prototypes, the long-tailed claim, missing implementation details, and resource overhead. The rebuttal addresses most of these points convincingly. In particular, the authors provide UMAP evidence for improved feature alignment, clarify that NIA normalizes per prototype rather than across all class prototypes, report rare-class gains, explain the prototype count and architecture settings, and show that the added communication and parameter overhead are small compared with FedAvg. Although the long-tailed framing was somewhat overstated, the authors acknowledge this and propose a more accurate title focused on mitigating client drift. Some limitations remain, especially the lack of stronger rare-class comparisons and more extensive statistical testing for the main table. Nevertheless, the method is technically sound, empirically effective, and relevant to heterogeneous federated medical learning. I therefore support acceptance, with the expectation that the final version includes the clarified positioning, overhead analysis, and limitations discussed in the rebuttal.
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 proposes FedProIn, a federated learning framework that decomposes client drift into feature drift and prototype drift, addressing them through Feature Divergence Loss (FDL) and Prototype Contrastive Loss (PCL) respectively, alongside a Normalized Influence Aggregation (NIA) mechanism for adaptive server-side aggregation.
Based on the reviewers’ feedback and the authors’ rebuttal, I recommend aceepting this manuscript. Two reviewers agree that their major concerns have been addressed, including the NIA normalization, hyperparameter sensitivity. And the authors should follow the suggestions to revise the final version and also adjust the long-tailed claim.