One Round Is All You Need: Analytic Federated Learning for Task-Heterogeneous Multi-Label Medical Image Classification

Afsaneh Mahanipour; Hana Khamfroush

REVIEW 3 major objections 3 minor 25 references

One Round Is All You Need: Analytic Federated Learning for Task-Heterogeneous Multi-Label Medical Image Classification

T0 review · 3 major / 3 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that task-heterogeneous federated multi-label medical image classification can be solved exactly with one communication round by aggregating per-class ridge-regression statistics from annotating clients.

desk verdict The per-class ridge aggregation theorem is correct; the headline beating-FedMLP numbers rest on a baseline run at one-tenth its required rounds. read the letter →

arxiv 2607.20641 v1 pith:6TBGZDFP submitted 2026-07-22 cs.LG

Afsaneh Mahanipour , Hana Khamfroush This is my paper

classification cs.LG

keywords federatedlearningtaskheterogeneitymulti-labelclassificationmedicalimageanalyticridgeregressionsufficientstatisticspseudo-labelrefinement

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that task heterogeneity in federated medical imaging—where each hospital labels only its own specialty's diseases—does not require iterative gradient training. It proposes a one-round analytic procedure in which each client uploads two sufficient statistics: a regularized feature autocorrelation matrix and one cross-correlation vector per class it annotates, computed with a balanced label projection that treats missing classes as absent rather than as negative. The server then solves one small linear system per disease class, and Theorem 1 guarantees the result is identical to the centralized balanced ridge-regression solution over all annotating clients' data. If the claim holds, federated training becomes a single closed-form aggregation step that is invariant to data partitioning, immune to client drift, and free of learning-rate hyperparameters, with strong performance even when every class is annotated by exactly one client.

What carries the argument

The per-class Absolute Aggregation law (Eq. 17) is the mechanism: it converts the federated task into a sum of sufficient statistics plus a single linear solve per class, cancelling the duplicated ridge regularizer γI so that the aggregate equals the centralized ridge solution. The balanced label projection (Eq. 9) is the companion mechanism that makes the objective well-posed under heterogeneity and imbalance, encoding missing classes as zero columns and normalizing positive and negative contributions to equal total mass without any reweighting hyperparameter.

What would settle it

Take a class with a single annotating client, hide the true labels of the other clients, and compare the server's one-round classifier against a centralized ridge classifier trained on the same balanced targets: if the two differ, the aggregation identity is wrong; if they match but a differently balanced objective (e.g., global prevalence targets) achieves higher BACC/AUC on the same test set, the balanced label projection is an inadequate surrogate.

Watch

Extended reading notes

Core claim

The central discovery is that the per-class Absolute Aggregation law (Eq. 17) reconstructs the exact centralized balanced ridge classifier. For each class c, the server forms A_c = Σ_{k∈S_c} A_k − (|S_c|−1)γI and b_c = Σ_{k∈S_c} b_{k,c}, then solves ŵ_c = A_c^{-1} b_c. Theorem 1 proves that this equals (H[c]^T H[c] + γI)^{-1} H[c]^T y[c], the unique minimizer of the centralized balanced ridge objective over the features and balanced target vectors of all clients that annotate class c. The companion balanced label projection replaces binary 0/1 entries with +1/N^+ and −1/N^− for active classes and zero for missing classes, so missing labels are structurally absent rather than false negatives

Load-bearing premise

The load-bearing premise is that minimizing a per-client balanced squared error on frozen features with target values +1/N+ and −1/N− produces a classifier that scores well under the reported clinical metrics (BACC, AUC, mAP); Theorem 1 guarantees exactness only for that constructed objective, and the paper does not show the two are aligned.

Editorial extensions

If this is right

Task-heterogeneous federated multi-label classification can be completed in one communication round with no gradient exchange; the two-round pseudo-label refinement is optional and in the reported experiments the one-round variant performs better.
Per-client communication is one d×d autocorrelation matrix plus one d-vector per annotated class, independent of the number of rounds.
In the extreme setting where each class is annotated by exactly one client, the one-round method retains 68.50% BACC and 74.92% AUC while gradient-based baselines drop to near chance-level performance.
Because the aggregation identity is purely algebraic, the global classifier is invariant to how data are partitioned across clients, and no learning rate, batch size, or local-step hyperparameters need tuning.
The method transfers across frozen backbones (ResNet, VGG, EfficientNet, and domain-pretrained DenseNet features) without backbone-specific adjustment.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper's exactness is exactness for the balanced squared-error objective, not for BACC or AUC; a reader should not infer that the +1/N+, −1/N− target encoding is the optimal surrogate for those metrics, and an alternative global-prevalence balancing scheme could change the reported gaps while preserving the algebraic guarantee.
The one-round-beats-two-rounds result implies pseudo-label refinement is not automatically beneficial when the first-round teacher is weak; a testable extension is to apply refinement only to classes with low annotator coverage or to tune the confidence threshold per class instead of fixing it at 0.7.
Because the aggregation law is linear, the same sufficient-statistics trick could be extended to multiple frozen feature extractors or to ridge-type heads on the same features, and could be stress-tested on feature distributions where one annotating client's data are not representative of the whole federation.
Theorem 1 is asserted without a proof in the text; the equality is a direct algebraic consequence of the ridge normal equations, and verifying it—or finding a counterexample—is the first check a skeptical reader would run.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Core take: the algebraic contribution is real and the empirical headline is shaky. Eq. (17) does exactly what the paper claims: aggregating the per-client regularized autocorrelations and balanced label projections over annotating clients reproduces the centralized ridge solution for that class. That is a clean extension of AFL to task-heterogeneous multi-label data, and the one-round protocol is a genuine communication-efficiency win if you live in the frozen-backbone linear-probe regime. The balanced label projection is a sensible way to keep rare classes from being crushed by negatives, and the hyperparameters are fixed, not tuned to hit the reported numbers.

What gave me pause is the comparison with FedMLP. The paper's own introduction says FedMLP requires 500 communication rounds, then Table I and the implementation section run it for 50. That is a tenfold under-training of the main competitor, and it shows: FedMLP sits near chance at Missing 7 (50.06 BACC), and that is where the 18.44 BACC gap comes from. If FedMLP is given its published budget, the gap may shrink substantially, and the 'consistently outperforms SOTA' claim is not yet supported. This is not a minor protocol detail; the paper's main empirical selling point depends on it. I'd ask for a rerun of all baselines at their intended round counts.

Two smaller soft spots. First, the optimality guarantee is for the balanced squared loss, not for BACC or AUC. The paper never argues the per-client-balanced surrogate is near-optimal for the clinical metrics. That is a modeling assumption, and it should be flagged as such. Second, there are no error bars or seeds, and the BACC uses a fixed 0.5 threshold, which can penalize calibration rather than discriminative power.

Credit where due: the paper is honest about its pseudo-label refinement round—it reports that single-round consistently beats two-round, and explains why. That admission is good, but it also means the pseudo-label mechanism should not be listed as a headline contribution in its current form.

Bottom line: this deserves a serious referee and, after a fair baseline comparison, might be a useful result for people who want exact ridge aggregation in one round. Send it to review with a requirement that the experiments be redone fairly and the contribution list reframed.

Referee Report

3 major / 3 minor

Summary. The paper proposes an analytic (gradient-free) federated learning method for multi-label medical image classification under task heterogeneity, where each client labels only a subset of the disease classes. The method replaces iterative training with three closed-form operations: a balanced label projection (Eq. 9), a per-class absolute aggregation law (Eqs. 15–17), and an optional confidence-thresholded pseudo-label refinement round (Eqs. 18–24). Theorem 1 states that the per-class aggregation recovers the centralized balanced ridge-regression solution over the annotating clients for that class. Experiments on ChestXray14 under four missing-class configurations report large gains over FedMLP and other baselines, e.g., up to +18.44 BACC and +13.24 AUC at Missing 7, while using at most two communication rounds.

Significance. The central algebraic contribution is real and valuable. Theorem 1, as verified by the reader, is correct: substituting the sufficient statistics A_k = H_k^T H_k + γI and b_{k,c} = H_k^T y_{k,c} into Eq. (17) reproduces the centralized balanced ridge solution for each class. If the empirical comparison is trustworthy, the paper solves a practically important FL task with an exact, one-round, parameter-light procedure, eliminating gradient hyperparameters and hundreds of communication rounds. The fixed hyperparameters in Table IV and the backbone-agnostic experiments in Table III are also strengths. However, the empirical claims currently rest on an under-converged FedMLP baseline, so the main significance cannot be assessed until that comparison is corrected.

major comments (3)

[§I, §IV.E, Table I, Table II] The paper explicitly states in Section I that FedMLP 'requires 500 communication rounds,' yet Table I lists Rounds=50 for FedMLP and Section IV.E trains all gradient-based baselines for R=50. FedMLP is a two-stage method whose warm-up and prototype pseudo-labeling phases are designed to improve over many rounds; under-running it by 10× is a plausible cause of its chance-level collapse at Missing 7 (50.06 BACC, 61.68 AUC), from which the headline gains are drawn. The claim of 'consistently outperforms the state-of-the-art' is therefore not supported by the current evidence. The authors should either run FedMLP and the other round-sensitive baselines for their published 500-round budget, or provide strong justification, with convergence curves, that 50 rounds is their converged operating point.
[Eq. (9), Theorem 1] Theorem 1 establishes optimality only with respect to the constructed per-client balanced ridge objective, not with respect to the reported evaluation metrics (BACC, AUC, mAP). In Eq. (9), positive labels are normalized by the local count N_{k,c}^+ and negatives by N_{k,c}^-; this balances each client's local contribution independently, so a class annotated by one large client and one small client is not globally balanced. The paper never demonstrates that the minimizer of this reweighted squared-error objective is near-optimal for BACC or AUC. This is a modeling assumption, not an algebraic error, but it is load-bearing for the empirical superiority claim. Please provide evidence (e.g., comparisons against a globally balanced encoding, or analysis of the relationship between the balanced ridge solution and BACC/AUC).
[Table IV, Eq. (25)] The inference threshold is fixed at 0.5 for all methods and all classes. BACC is threshold-dependent (Eq. 25), and a fixed 0.5 threshold can penalize a model whose scores are not calibrated to 0.5, conflating calibration with discriminative power. Since BACC is the primary metric, the comparison should either use per-class threshold selection, report threshold-free metrics as the primary evidence, or justify the fixed threshold as methodologically neutral. This is secondary to the baseline-convergence issue but still affects the headline numbers.

minor comments (3)

[§IV.F] The sentence attributing FedMLP's collapse to 'prototype-guided pseudo-labeling' becoming unreliable with one annotating client is a causal claim not directly tested by the experiments; consider softening or adding an ablation that isolates this mechanism.
[Table IV] The pseudo-label hyperparameters τ, α, δ+, and δ- are fixed without sensitivity analysis. Since the two-round variant underperforms the one-round variant in all configurations, a sensitivity study would clarify whether the chosen values are responsible and whether the refinement stage deserves its status as a contribution.
[§II.A] The background description of FedAvg and its variants is standard but somewhat longer than needed; the derivation in Section III is the core of the paper.

Circularity Check

0 steps flagged · score 1.0 of 10

Central ridge-aggregation identity is algebraic and externally benchmarked; no circular derivation.

full rationale

The paper's core derivation is self-contained linear algebra rather than a circular fit. Equation (17) is constructed directly from the local sufficient statistics defined in Eqs. (13)-(16): substituting A_c = sum_{k in S_c} H_k^T H_k + γI and b_c = sum_{k in S_c} H_k^T y_{k,c} into w_hat_c = A_c^{-1} b_c yields exactly (H[c]^T H[c] + γI)^{-1} H[c]^T y[c]. Theorem 1 therefore certifies an algebraic identity, not a fitted prediction. The balanced label projection in Eq. (9) is an a priori modeling choice; the theorem only claims optimality for that constructed ridge objective, and the paper does not claim that this objective is globally optimal for BACC or AUC. Table IV fixes all hyperparameters (γ=1, τ=0.7, α=0.5, δ+=5, δ-=50) rather than fitting them to reproduced numbers. The base absolute-aggregation law is credited to the external AFL work [14], and the only self-references ([3], [15]) are contextual and do not carry the derivation. The notable empirical risk—that FedMLP is run for 50 rounds even though the manuscript itself says FedMLP requires 500 rounds—is a baseline-fairness and correctness concern, not a circular reduction, so it does not raise the circularity score. Overall, no load-bearing step reduces to its own inputs or to a self-citation chain.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The primary 1-round claim rests on the frozen-shared-backbone domain assumption, the per-client balanced-label surrogate objective, and the disjointness/coverage assumptions; plus one relevant hand-set constant (γ=1.0) and the fixed inference threshold. Round-2 hyperparameters (τ, α, δ±) are free but not load-bearing since Round 2 is reported as inferior.

free parameters (5)

Ridge regularization γ = 1.0
Table IV; fixed shrinkage/invertibility constant in every A_k and in the server solve (Eqs. 12, 15, 17); chosen by hand, no validation protocol stated.
Pseudo-label confidence threshold τ = 0.7
Table IV; controls acceptance of pseudo-positives/negatives in Eq. (19); only affects Round 2, which the paper reports as inferior to Round 1.
Pseudo-label weight α = 0.5
Table IV; blends real and pseudo label projections in Eq. (23); only affects Round 2.
Minimum pseudo-counts δ+ / δ− = δ+=5, δ−=50
Table IV; filters clients in Eq. (20); only affects Round 2.
Inference threshold = 0.5
Table IV; BACC (Eq. 25) is computed at this fixed threshold for all methods, which materially decides the headline BACC comparisons.

assumptions (6)

domain assumption All clients share one frozen pre-trained backbone f_backbone(·, θ); features are extracted once and never updated (Eq. 7).
The method is a linear probe on fixed features; the closed-form solve is optimal only for that feature space. Performance ceiling and cross-backbone variation (Table III) show how load-bearing this is.
domain assumption The balanced label encoding of Eq. (9) is an appropriate surrogate objective for the reported metrics (BACC/AUC).
Theorem 1 proves optimality only for the constructed balanced squared loss, not for BACC/AUC; the empirical story depends on this unproven alignment. Also the balancing is per client, not global.
domain assumption Every class is annotated by at least one client (S_c ≠ ∅).
Required to form A_c and b_c in Eqs. (15)-(16); stated in Section III.A.
domain assumption Client datasets are disjoint and missing labels are structurally absent, so vertical concatenation in Theorem 1 corresponds to the federation.
Theorem 1's equivalence uses H[c] = [H_{k1};...;H_{kKc}]. The experimental section never describes the image-level partition into 8 clients.
standard math Standard linear algebra for ridge regression normal equations and block sums (Σ_k H_k^T H_k = H^T H for concatenated H).
Used in Theorem 1 and Eqs. (15)-(17); standard and machine-checkable.
ad hoc to paper Sigmoid-mapped scores with thresholds τ and 1−τ are meaningful confidence estimates for pseudo-labeling (Eqs. 18-19).
Heuristic confidence filtering; only affects Round 2, reported as inferior, so this axiom is not load-bearing for the primary claim.

how reviews work

0 comments

Cite this review

Pith. "Pith review of One Round Is All You Need: Analytic Federated Learning for Task-Heterogeneous Multi-Label Medical Image Classification." pith.science (2026). https://pith.science/paper/6TBGZDFP

@misc{pith2026260720641,
  author       = {Pith},
  title        = {Pith review of: One Round Is All You Need: Analytic Federated Learning for Task-Heterogeneous Multi-Label Medical Image Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6TBGZDFP}},
  note         = {Machine review of arXiv:2607.20641}
}

read the original abstract

Federated learning (FL) enables multiple clinical institutions to collaboratively train a shared disease classifier without centralizing patient data. In practice, however, each institution annotates only the pathologies within its area of expertise, so the federation operates under task heterogeneity: each client holds labels for a strict subset of the target disease categories while the remaining classes are entirely unobserved at that site. Existing gradient-based FL methods fail under this setting because they require hundreds of communication rounds to converge and because missing class labels introduce systematic false-negative bias that the model cannot correct without a principled mechanism. We propose an analytic federated learning framework for multi-label medical image classification under task heterogeneity. The proposed method replaces iterative gradient optimization with three closed-form operations: a balanced label projection that neutralizes class-imbalance bias by normalizing positive and negative contributions to equal total mass; a per-class absolute aggregation law that independently assembles the optimal ridge-regression classifier for each disease category from the sufficient statistics uploaded by its annotating clients; and an optional analytic pseudo-label refinement round that propagates missing-class knowledge from a confidence-filtered teacher classifier to non-annotating clients. The entire procedure requires at most two communication rounds, irrespective of the degree of task heterogeneity or the number of participating clients. Experiments on ChestXray14 under four progressively severe missing-class configurations demonstrate that the proposed method consistently outperforms the state-of-the-art federated multi-label method FedMLP by up to 18.44 BACC points and 13.24 AUC points, while reducing the communication.

Figures

Figures reproduced from arXiv: 2607.20641 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 3 linked inside Pith

[1]

Health insurance portability and accountability act,

P. F. Edemekong, P. Annamaraju, and M. J. Haydel, “Health insurance portability and accountability act,” 2018

2018
[2]

Communication-efficient learning of deep networks from decentralized data,

B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” inArtificial intelligence and statistics. Pmlr, 2017, pp. 1273–1282

2017
[3]

Feder- ated reprogramming knowledge distillation for medical image classification,

A. Mahanipour, A.-A.-Z. Imran, and H. Khamfroush, “Feder- ated reprogramming knowledge distillation for medical image classification,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2025, pp. 143–152

2025
[4]

Probing the efficacy of federated parameter-efficient fine-tuning of vision transformers for medical image classifica- tion,

N. Alkhunaizi, F. Almalik, R. Al-Refai, M. Naseer, and K. Nan- dakumar, “Probing the efficacy of federated parameter-efficient fine-tuning of vision transformers for medical image classifica- tion,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2024, pp. 236– 245

2024
[5]

Fedmlp: Federated multi-label medical image classification under task heterogeneity,

Z. Sun, N. Wu, J. Shi, L. Yu, K.-T. Cheng, and Z. Yan, “Fedmlp: Federated multi-label medical image classification under task heterogeneity,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2024, pp. 394–404

2024
[6]

Rscfed: Random sampling consensus federated semi-supervised learning,

X. Liang, Y . Lin, H. Fu, L. Zhu, and X. Li, “Rscfed: Random sampling consensus federated semi-supervised learning,” inPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 154–10 163

2022
[7]

Fixmatch: Simplifying semi-supervised learning with consistency and con- fidence,

K. Sohn, D. Berthelot, N. Carlini, Z. Zhang, H. Zhang, C. A. Raffel, E. D. Cubuk, A. Kurakin, and C.-L. Li, “Fixmatch: Simplifying semi-supervised learning with consistency and con- fidence,”Advances in neural information processing systems, vol. 33, pp. 596–608, 2020

2020
[8]

Federated semi- supervised medical image classification via inter-client relation matching,

Q. Liu, H. Yang, Q. Dou, and P.-A. Heng, “Federated semi- supervised medical image classification via inter-client relation matching,” inInternational conference on medical image com- puting and computer-assisted intervention. Springer, 2021, pp. 325–335

2021

Show all 25 references

[9]

Fed- noro: Towards noise-robust federated learning by addressing class imbalance and label noise heterogeneity,

N. Wu, L. Yu, X. Jiang, K.-T. Cheng, and Z. Yan, “Fed- noro: Towards noise-robust federated learning by addressing class imbalance and label noise heterogeneity,”arXiv preprint arXiv:2305.05230, 2023

2023 arXiv
[10]

Towards federated learning against noisy labels via local self-regularization,

X. Jiang, S. Sun, Y . Wang, and M. Liu, “Towards federated learning against noisy labels via local self-regularization,” in Proceedings of the 31st ACM International Conference on In- formation & Knowledge Management, 2022, pp. 862–873

2022
[11]

Learning from the kernel and the range space,

K.-A. Toh, “Learning from the kernel and the range space,” in 2018 IEEE/ACIS 17th International Conference on Computer and Information Science (ICIS). IEEE, 2018, pp. 1–6

2018
[12]

Blockwise recursive moore– penrose inverse for network learning,

H. Zhuang, Z. Lin, and K.-A. Toh, “Blockwise recursive moore– penrose inverse for network learning,”IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 52, no. 5, pp. 3237– 3250, 2021

2021
[13]

Acil: Analytic class-incremental learning with absolute memo- rization and privacy protection,

H. Zhuang, Z. Weng, H. Wei, R. Xie, K.-A. Toh, and Z. Lin, “Acil: Analytic class-incremental learning with absolute memo- rization and privacy protection,”Advances in Neural Information Processing Systems, vol. 35, pp. 11 602–11 614, 2022

2022
[14]

Afl: A single-round analytic approach for federated learning with pre-trained models,

R. He, K. Tong, D. Fang, H. Sun, Z. Zeng, H. Li, T. Chen, and H. Zhuang, “Afl: A single-round analytic approach for federated learning with pre-trained models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 4988–4998

2025
[15]

Embedded federated feature selection with dynamic sparse training: balancing accuracy-cost tradeoffs,

A. Mahanipour and H. Khamfroush, “Embedded federated feature selection with dynamic sparse training: balancing accuracy-cost tradeoffs,” in2025 International Joint Conference on Neural Networks (IJCNN). IEEE, 2025, pp. 1–8

2025
[16]

Federated optimization in heterogeneous networks,

T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, and V . Smith, “Federated optimization in heterogeneous networks,” Proceedings of Machine learning and systems, vol. 2, pp. 429– 450, 2020

2020
[17]

Tackling the objective inconsistency problem in heterogeneous federated op- timization,

J. Wang, Q. Liu, H. Liang, G. Joshi, and H. V . Poor, “Tackling the objective inconsistency problem in heterogeneous federated op- timization,”Advances in neural information processing systems, vol. 33, pp. 7611–7623, 2020

2020
[18]

Federated learning based on dynamic regularization,

D. A. E. Acar, Y . Zhao, R. M. Navarro, M. Mattina, P. N. Whatmough, and V . Saligrama, “Federated learning based on dynamic regularization,”arXiv preprint arXiv:2111.04263, 2021

2021 arXiv
[19]

A theory of networks for approximation and learning,

T. Poggio and F. Girosi, “A theory of networks for approximation and learning,” Tech. Rep., 1989

1989
[20]

Chestx-ray8: Hospital-scale chest x-ray database and bench- marks on weakly-supervised classification and localization of common thorax diseases,

X. Wang, Y . Peng, L. Lu, Z. Lu, M. Bagheri, and R. M. Summers, “Chestx-ray8: Hospital-scale chest x-ray database and bench- marks on weakly-supervised classification and localization of common thorax diseases,” inProceedings of the IEEE conference on computer vision and patte...

2017
[21]

Class balanced adaptive pseudo labeling for federated semi-supervised learning,

M. Li, Q. Li, and Y . Wang, “Class balanced adaptive pseudo labeling for federated semi-supervised learning,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 16 292–16 301

2023
[22]

Torchxrayvision: A library of chest x-ray datasets and models,

J. P. Cohen, J. D. Viviano, P. Bertin, P. Morrison, P. Torabian, M. Guarrera, M. P. Lungren, A. Chaudhari, R. Brooks, M. Hashir et al., “Torchxrayvision: A library of chest x-ray datasets and models,” inInternational Conference on Medical Imaging with Deep Learning. PMLR, 2022...

2022
[23]

Deep residual learning for image recognition,

K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

2016
[24]

Very deep convolutional networks for large-scale image recognition,

K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,”arXiv preprint arXiv:1409.1556, 2014

2014 arXiv
[25]

Efficientnet: Rethinking model scaling for convolutional neural networks,

M. Tan and Q. Le, “Efficientnet: Rethinking model scaling for convolutional neural networks,” inInternational conference on machine learning. PMLR, 2019, pp. 6105–6114

2019

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

Health insurance portability and accountability act,

P. F. Edemekong, P. Annamaraju, and M. J. Haydel, “Health insurance portability and accountability act,” 2018

2018

[2] [2]

Communication-efficient learning of deep networks from decentralized data,

B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” inArtificial intelligence and statistics. Pmlr, 2017, pp. 1273–1282

2017

[3] [3]

Feder- ated reprogramming knowledge distillation for medical image classification,

A. Mahanipour, A.-A.-Z. Imran, and H. Khamfroush, “Feder- ated reprogramming knowledge distillation for medical image classification,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2025, pp. 143–152

2025

[4] [4]

Probing the efficacy of federated parameter-efficient fine-tuning of vision transformers for medical image classifica- tion,

N. Alkhunaizi, F. Almalik, R. Al-Refai, M. Naseer, and K. Nan- dakumar, “Probing the efficacy of federated parameter-efficient fine-tuning of vision transformers for medical image classifica- tion,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2024, pp. 236– 245

2024

[5] [5]

Fedmlp: Federated multi-label medical image classification under task heterogeneity,

Z. Sun, N. Wu, J. Shi, L. Yu, K.-T. Cheng, and Z. Yan, “Fedmlp: Federated multi-label medical image classification under task heterogeneity,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2024, pp. 394–404

2024

[6] [6]

Rscfed: Random sampling consensus federated semi-supervised learning,

X. Liang, Y . Lin, H. Fu, L. Zhu, and X. Li, “Rscfed: Random sampling consensus federated semi-supervised learning,” inPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 154–10 163

2022

[7] [7]

Fixmatch: Simplifying semi-supervised learning with consistency and con- fidence,

K. Sohn, D. Berthelot, N. Carlini, Z. Zhang, H. Zhang, C. A. Raffel, E. D. Cubuk, A. Kurakin, and C.-L. Li, “Fixmatch: Simplifying semi-supervised learning with consistency and con- fidence,”Advances in neural information processing systems, vol. 33, pp. 596–608, 2020

2020

[8] [8]

Federated semi- supervised medical image classification via inter-client relation matching,

Q. Liu, H. Yang, Q. Dou, and P.-A. Heng, “Federated semi- supervised medical image classification via inter-client relation matching,” inInternational conference on medical image com- puting and computer-assisted intervention. Springer, 2021, pp. 325–335

2021