PaperlayerView exact revision on arXiv

Source-linked AI summary

FedDRAW: Federated Dual Reputation Annealing Weighting for Heterogeneous Multi-Institutional Chest Radiograph Classification

Maryam Moradpour, Anne-Christin Hauschild

arXiv:2609.05223v1cs.LG

TL;DR

The paper addresses the limitation of sample-size-proportional aggregation in heterogeneous federated medical imaging, where informative small clients can receive little influence. FedDRAW combines data size and client-to-global model similarity through inner and outer annealing schedules, and across twelve CheXpert and ChestMNIST partition scenarios it achieved the highest average rank among eight methods on both AUC and GM.

  • Problem

    Sample-size-proportional aggregation can give large clients disproportionate influence even when smaller clients hold informative or complementary institutional knowledge.

  • Method

    FedDRAW is a server-side method that combines client data size and client-to-global model similarity under coupled inner and outer annealing schedules.

  • Results

    FedDRAW achieved the highest GM in all seven CheXpert configurations, the highest AUC in five, and the highest values of both metrics in principal heterogeneous ChestMNIST configurations.

  • Takeaways & Limitations

    Scheduling aggregation signals across training provides a stage-aware alternative to fixing client influence by sample count alone.

  • Takeaways & Limitations

    The reported AUC and related metrics respond weakly to class imbalance and do not establish clinically acceptable performance at a chosen operating threshold.

Abstract

from arXiv · show

Artificial intelligence models are promising for medical diagnosis, but they require large numbers of unbiased data, which in medicine are distributed across hospitals and cannot be centralized to protect patient privacy. Federated Learning (FL) addresses this, since hospitals train one shared diagnostic model while patient data remain local. Training proceeds in communication rounds, in which each hospital trains the shared model locally and returns it to the server for merging by weighted average. This aggregation weight determines whose institutional knowledge shapes the result. Federated averaging (FedAvg) sets it in proportion to local sample count, so a small but informative hospital is permanently assigned a small influence, andl argest clients could dominate the global model even when they are less informative. We propose Federated Dual Reputation Annealing Weighting (FedDRAW), a server-side aggregation method that combines a data-size prior with the cosine similarity between client and global parameters under two coupled annealing schedules. An inner schedule shifts client reputation from the size prior towards similarity. An outer, deferred annealing schedule on the softmax inverse temperature keeps the weighting selective in the early and middle rounds and relaxes it to uniformity at convergence. We evaluate FedDRAW on 12 simulated client-partition scenarios of two chest radiograph datasets (CheXpert and ChestMNIST), against seven federated baselines under identical local training settings. FedDRAW achieved the highest average rank among all eight methods under both AUC and the geometric mean (GM) of sensitivity and specificity, which a Friedman test with Nemenyi post-hoc analysis confirmed to be a statistically significant difference between the methods. Scheduling two signals, rather than fixing the weights by sample count alone, could enable less biased diagnostic models.

1. Introduction

The paper identifies shortcomings in sample-size-only aggregation for heterogeneous federated medical imaging and proposes FedDRAW, which schedules data-size and model-similarity signals across training while moving toward equal treatment.

  • Single-institution medical datasets may not represent diverse patient populations, scanners, protocols, or annotation practices, motivating collaborative federated training.
  • FedAvg weights client models by local sample count, which can restrict small but informative institutions and amplify poorly trained or distributionally biased updates.
  • Existing approaches use model similarity, label-distribution discrepancy, or validation performance, but generally apply their weighting criteria without deliberately scheduling them with training progress.
  • FedDRAW combines client data size with client-to-global model similarity through an inner schedule that gradually shifts reliance toward similarity as the global reference matures.
  • FedDRAW is evaluated on twelve heterogeneous partitions of CheXpert and ChestMNIST against seven aggregation baselines under identical local hyperparameters.
  • A second, outer temperature schedule relaxes aggregation differences toward uniformity in later rounds, defining aggregation-level equal-treatment fairness.

2. Related Work

Related work spans local-optimization methods, aggregation rules beyond sample count, fairness objectives, and medical-imaging evaluation concerns; FedDRAW is positioned as stage-aware aggregation under clinical data heterogeneity.

  • Federated Optimization under Heterogeneous Client Data: Federated optimization methods address heterogeneous client objectives by constraining local deviation, correcting client drift, or normalizing unequal local computation.
  • Aggregation Weighting beyond Sample Count: Aggregation-focused methods use label-distribution discrepancy, parameter distance, gradient alignment, or server-side validation performance instead of relying solely on local sample count.
  • Aggregation Weighting beyond Sample Count: Existing aggregation formulations may change weights as client states change, but their scoring or optimization criterion is generally applied identically in every round rather than deliberately scheduled.
  • Fairness: Federated-learning fairness includes performance, demographic, contribution, and aggregation-level notions that equalize different quantities.
  • Federated Learning in Medical Imaging: Medical-imaging federations combine institutional shifts in hardware, protocols, populations, prevalence, and annotation practices, making volume-only weighting insufficient to represent diversity.
  • Federated Learning in Medical Imaging: AUC and related metrics respond weakly to class imbalance, while threshold-dependent clinical decisions involve unequal consequences for false positives and false negatives.
  • Federated Learning in Medical Imaging: The study therefore complements macro-averaged AUC with the geometric mean of sensitivity and specificity using validation-selected, fixed per-class thresholds.

3. Method

The method section frames FedDRAW as a server-side framework for federated chest-radiograph classification built around two coupled annealing mechanisms that separately govern client reputation and weight concentration.

  • FedDRAW is presented as a server-side aggregation framework for experiments on two public multi-label chest-radiograph datasets.
  • Its inner annealing controls the informational basis of client reputation, while its outer annealing controls how strongly reputation differences become aggregation-weight differences.

3.1. Datasets

The study evaluates a consistent five-finding multi-label task across CheXpert and ChestMNIST, which differ in scale, resolution, populations, and label-generation pipelines.

  • The experiments use CheXpert and ChestMNIST to test whether aggregation behavior remains consistent across different data sources and annotation pipelines.
  • Label distributions: Table 1 reports positive-image counts for each finding in CheXpert and ChestMNIST, with counts not mutually exclusive because images may have multiple labels.
  • CheXpert: CheXpert contains 224,316 chest radiographs from 65,240 patients, with five selected findings converted to binary labels after uncertainty handling.
  • Example images and labels: Figure 1 shows representative test images and binary labels, using radiologist majority-vote ground truth for CheXpert and automated source-report labels for ChestMNIST.
  • ChestMNIST: ChestMNIST derives from ChestX-ray14, retains the same five pathologies, and supplies binary labels generated through automated report text mining.
  • ChestMNIST: ChestMNIST retains substantial class imbalance and lower image resolution, providing a complementary and more challenging counterpart to CheXpert.

3.2. Preliminaries and Problem Formulation

Federated learning enables hospitals to train a shared model without exchanging raw patient data, but heterogeneous cohorts make aggregation challenging. FedAvg uses static sample-count weights, motivating a method that adapts client influence over training.

  • Federated learning coordinates a shared model across private institutional datasets without exchanging raw patient data.
  • Horizontal cross-silo FL lets institutions train identical architectures whose parameters can be compared and merged at the server.
  • FedAvg fixes each client’s weight to its sample fraction throughout training, assuming dataset size sufficiently represents contribution quality.
  • Static size-based weighting can marginalize small but informative clients and cannot adapt to changing client quality during training.
  • FedDRAW addresses stale size weights and unreliable early similarity by transitioning from a stabilizing size prior toward similarity while preserving client participation.

3.3. Overview

FedDRAW uses two server-side annealing schedules to determine client influence. Inner annealing defines reputation from data size and model similarity, while outer annealing converts reputation differences into aggregation weights.

  • FedDRAW jointly anneals client reputation and the strength with which reputation differences determine aggregation weights.
  • Inner annealing progressively shifts reputation from a static data-size prior toward dynamic similarity with the global model.
  • Outer annealing uses a temperature-controlled softmax to convert reputations into normalized aggregation weights.
  • Both schedules vary across communication rounds, moving aggregation from cautious size stabilization toward similarity-informed weighting and eventual fairness.

3.4. Algorithm

FedDRAW preserves the standard federated training workflow while performing reputation and weight calculations on the server from uploaded parameters. It adds no extra server model or communication beyond standard parameter uploads.

  • FedDRAW adds no server-side model and requires no client-to-server communication beyond standard parameter uploads.
  • The server computes all reputation and aggregation weights from parameters it already receives.
  • Clients train locally from the broadcast global model for E epochs in parallel at every communication round.

3.5. Inner Annealing: Reputation Scoring

The inner annealing stage gradually changes the reputation criterion from client data size to model similarity. An exponentially decaying control provides early size-based stabilization and later similarity-based refinement.

  • Inner annealing initially uses client data size and progressively transitions toward similarity between updated client and global models.
  • Using similarity alone from the start risks cold-start effects because the immature global model provides an unreliable reference.
  • The size prior is static, while cosine similarity is recomputed each round as client and global parameters evolve.
  • The reputation score is a convex combination controlled by γ_t, with γ_t = 1 giving pure size-based reputation and γ_t = 0 giving pure similarity.
  • The exponential schedule γ_t = exp(−ηt) starts at 1, decays smoothly toward 0, and shifts most rapidly during early rounds.
  • The annealing rate η controls transition speed: larger values suit faster convergence, whereas smaller values retain size-based stabilization longer.

3.6. Outer Annealing: From Selective to Fair Aggregation

FedDRAW’s outer annealing controls how reputation differences become aggregation-weight differences, preserving selective aggregation early and moving toward fairness through near-uniform weighting late in training.

  • Fairness objective: Client-level fairness is operationalized as preventing any client from being permanently excluded from meaningful participation in the global model.
  • Selective-to-fair transition: The outer stage controls how strongly reputation differences translate into weight differences and decreases this strength over time.Early selectivity favors larger, more representative datasets; late equalization prevents weaker or dissimilar clients from remaining marginalized.
  • Weight conversion: The softmax inverse temperature β_t determines how sharply reputation scores become normalized aggregation weights.Large β_t amplifies small reputation differences, whereas β_t → 0 yields equal weighting.
  • Deferred annealing: The deferred schedule keeps β_t near β_max through early and middle rounds, then reaches β_T = 0 so final aggregation weights are exactly uniform.This postpones flattening until the final rounds and avoids premature redistribution that could discard small clients’ contributions.
  • Coupled schedules: The inner schedule shifts reputation from the data-size prior toward similarity before the outer schedule equalizes weights.The two schedules therefore separate what defines reputation from how strongly reputation affects aggregation.

3.7. Weighted Aggregation

FedDRAW couples inner reputation annealing with outer weight annealing while retaining the standard federated-averaging update, making its novelty the dynamic computation of aggregation weights.

  • Dual annealing process: At each round, the inner stage computes client reputations while the outer stage computes aggregation weights on separate decay schedules.The inner parameter controls what defines reputation, whereas the outer parameter controls how sharply reputation differences become weight differences.
  • Schedule interaction: The coupled process begins with size-driven selective aggregation and progresses toward similarity-informed, more equal treatment across clients.The inner scorer shifts from data size toward similarity while the outer supervisor relaxes toward equal treatment.
  • Design rationale: Only joint inner and outer annealing achieves changing reputation content, selective early aggregation, and late-round fairness together.Using only outer annealing leaves reputation based on one static signal, while using only inner annealing can suppress low-similarity clients late.
  • Aggregation update: FedDRAW uses the standard federated-averaging update, with novelty confined to dynamically computing round-dependent aggregation weights.

3.8. Last-Layer Similarity

FedDRAW computes client similarity from the final classification layer because it is more sensitive to client-specific label-distribution differences than generic early representations.

  • Similarity representation: FedDRAW computes cosine similarity using only final-layer parameters rather than the full model vector.
  • Why not the full model: Shared initialization and few local epochs leave early generic layers similar across clients, causing whole-model similarity to dilute client-specific learning signals.Whole-model similarity can therefore produce artificially high values and insufficiently capture relationships between clients.
  • Last-layer rationale: The final layer maps representations to predictions and is most directly tied to local label distributions, making it especially sensitive to client data-distribution skew.

4. Experimental setup and evaluation

FedDRAW was evaluated on heterogeneous client partitions of CheXpert and ChestMNIST using AUC and GM, with comparisons against FedAvg and other federated methods. Across these scenarios, it consistently improved GM and often improved AUC, including settings with severe quantity or class-distribution imbalance.

  • Metrics: AUC and GM were macro-averaged over five pathologies, with GM complementing AUC by evaluating sensitivity and specificity together at deployed thresholds.GM is the square root of sensitivity multiplied by specificity and is less tolerant of majority-class bias.
  • Client partition scenarios: The experiments constructed controlled client partitions because the datasets lacked predefined institutions with differing sizes or pathology distributions.CheXpert varied quantity imbalance and enriched smaller clients with pathology-specific samples; ChestMNIST used increasingly severe imbalance and Atelectasis enrichment.
  • CheXpert results: FedDRAW achieved the highest GM in every CheXpert scenario and the highest AUC in five of seven scenarios.In the other two CheXpert scenarios, its AUC was within 0.19 and 0.16 percentage points of the best value.
  • CheXpert results: In CHX-1, FedDRAW reached 86.67 ± 0.25% AUC and 78.76 ± 0.18% GM, versus 84.98 ± 0.68% and 76.01 ± 1.31% for FedAvg.These corresponded to improvements of 1.69 and 2.75 percentage points, respectively.
  • CheXpert results: FedDRAW improved over FedAvg in CHX-2 and CHX-3 by 1.69 and 2.42 percentage points, and by 0.93 and 1.32 percentage points, for AUC and GM respectively.The reported FedDRAW values were 87.46 ± 0.30% AUC and 77.83 ± 0.81% GM in CHX-2, and 86.64 ± 0.18% AUC and 78.03 ± 0.91% GM in CHX-3.
  • CheXpert results: FedDRAW also outperformed FedAvg in class-enriched CheXpert scenarios, with AUC gains of 1.55 and 2.38 percentage points and GM gains of 2.39 and 1.70 percentage points.No competing method exceeded FedDRAW in either metric in CHX-5 or CHX-6.
  • ChestMNIST results: On ChestMNIST, FedDRAW achieved the highest GM in four heterogeneous configurations, while AUC differences between leading methods were small.In CHM-1 and CHM-2, FedDRAW improved over FedAvg by 1.65 and 2.77 percentage points in AUC, and by 1.78 and 2.64 percentage points in GM.
  • ChestMNIST results: In CHM-4, FedDRAW reached 82.67 ± 0.77% AUC and 75.48 ± 0.42% GM, compared with 81.69 ± 0.21% and 73.67 ± 0.41% for FedAvg.The corresponding improvements were 0.98 and 1.81 percentage points, and the result was consistent with the CheXpert class-enriched experiments.

5. Conclusion

FedDRAW replaces fixed sample-size aggregation with annealed weighting that shifts from data-size priors toward model similarity and ultimately uniformity. Across heterogeneous chest-radiograph partitions, it achieved strong performance, while remaining limited to one scalar weight per client.

  • 5. Conclusion: FedDRAW blends client–global parameter cosine similarity with a data-size prior through two coupled exponential annealing schedules.The size prior anchors early aggregation, similarity gains influence during training, and decaying temperature drives weights toward uniformity at convergence.
  • 5. Conclusion: FedDRAW achieved the highest GM in all seven CheXpert configurations and the highest AUC in five, while also leading both metrics in principal heterogeneous ChestMNIST configurations.Its largest gains occurred when many clients were small or class-enriched, where proportional weighting was most restrictive.
  • 5. Conclusion: FedDRAW assigns one scalar weight per client, so a client informative for one pathology but not others cannot receive pathology-specific influence.The authors identify class-wise aggregation weights as a future extension.
Loading 2609.05223v1…