Swarm of digital particles trains medical AI without sharing patient data

Every hospital holds only a fragment of the diagnostic puzzle. One institution may archive thousands of lung and colon histopathology slides, another may own a few hundred immunohistochemical breast tissue images, and a third may keep microscopic blood smears that expose leukemia cell by cell. Privacy laws such as the General Data Protection Regulation and the Health Insurance Portability and Accountability Act forbid pooling these archives, so medical AI has increasingly turned to federated learning, a decentralized training scheme in which a shared model travels to each clinic, learns on local data, and returns only parameter updates. Yet a study published in Cluster Computing argues that the way most federated pipelines choose which hospitals participate in each training round is quietly squandering the approach’s potential. A Brazilian research team led by Leonardo Gabriel Ferreira Rodrigues, Larissa Ferreira Rodrigues Moreira, Rodrigo Moreira and André Ricardo Backes replaced random participant sampling with particle swarm optimization, a nature-inspired search algorithm that hunts for the most informative institutions round after round.

The obstacle the team targeted is statistical heterogeneity. In federated learning, the data held by each client are rarely independent and identically distributed. Hospitals differ in imaging equipment, staining protocols, patient demographics, disease prevalence and annotation quality, which means the standard Federated Averaging algorithm, or FedAvg, operates on a federation of unequal contributors. When participants are chosen at random, institutions whose data contribute little may dominate the updates while clinics holding rare but diagnostically valuable samples sit idle. Choosing the best subset of k clients from a pool of n candidates to maximize a global utility is formally classified as NP-hard, a combinatorial problem whose exact solutions become computationally intractable as the client count grows. In medical imaging this difficulty is amplified, because the inclusion of the right participants can be decisive for model sensitivity to underrepresented classes such as rare cancer subtypes or uncommon morphological patterns.

The team’s answer borrows from the collective behavior of bird flocks and fish schools. Particle Swarm Optimization, or PSO, dispatches a population of particles through a search space, where each particle maintains a position and a velocity. At every iteration the velocity is updated by an equation that blends three influences: an inertia weight that preserves the previous direction, a cognitive term pulling the particle toward the best position it has found so far, and a social term pulling it toward the best position discovered by any particle in the swarm, with random coefficients keeping the search exploratory. In the proposed formulation, each particle encodes a continuous vector whose components act as selection priorities for the available clients, and the algorithm maps this continuous encoding to a discrete subset by picking the k clients with the highest priorities. The fitness of each candidate is simply the sum of the performance scores of the clients it selects.

Those scores are empirical rather than modeled. After a client completes local training and evaluates the resulting model on its own test partition, the server records the reported accuracy as the client’s utility, refreshed whenever the client participates again. Until enough scores have accumulated, the system defaults to random selection. Crucially, the optimizer touches neither local training nor the aggregation mechanism: FedAvg remains fully intact, and PSO acts purely as a data-driven orchestration layer that decides who trains in each round. The researchers ran the swarm with fixed hyperparameters drawn from established swarm intelligence guidelines, using 20 particles, symmetric cognitive and social coefficients of 1.8, an inertia weight of 0.7 and a budget of 500 fitness evaluations per round, keeping the selection step light enough that it never becomes a bottleneck in a federated pipeline.

The evaluation spanned three public medical imaging datasets chosen to represent a spectrum of federated realities. LC25000 contains 25,000 histopathological images of 768 by 768 pixels covering colon and lung tissue, merged into binary normal versus cancer tasks. Biglycan comprises 336 immunohistochemical photomicrographs of breast tissue from the pathology archives of Hospital de Clínicas de Porto Alegre in Brazil, imbalanced between 203 cancer and 133 healthy images, reflecting the scarcity of single-institution archives. ALL-IDB2 offers microscopic single-cell blood smear images for classifying Acute Lymphoblastic Leukemia. Four convolutional neural network architectures, SqueezeNet, MobileNet-V2, ResNet-18 and EfficientNet-B0, initialized with ImageNet pre-trained weights, were trained under heterogeneous class-based and Dirichlet non-IID partitions on the Fabric testbed, equipped with an NVIDIA Quadro RTX 6000 GPU and the Flower federated framework. The proposed strategy was benchmarked against FedAvg with random selection, the utility-guided Oort framework, the fine-grained PyramidFL method and a Genetic Algorithm baseline.

The headline finding is that a lightweight swarm matches a heavyweight contender. Across 28 experimental configurations, the Friedman test confirmed highly significant differences among the strategies, and pairwise Wilcoxon signed-rank tests with Bonferroni correction showed that PSO is statistically indistinguishable from Oort: the two methods won 13 and 14 configurations respectively with one tie, an adjusted p-value of 1.00 and a negligible Cliff’s delta of minus 0.074. This equivalence is striking because Oort requires explicit modeling of client utility through loss gradients and training duration, whereas the PSO selector needs nothing more than empirical local accuracy. Both approaches decisively outperformed PyramidFL, with PSO posting a large effect size of plus 0.80 and mean accuracy improvements over PyramidFL ranging from 5.8 percentage points on SqueezeNet to 22.6 percentage points on MobileNet. The F1-score analysis reproduced every one of these patterns.

The benefits, however, are architecture-dependent. Lightweight networks proved the most sensitive to the quality of the selected participants: MobileNet gained more than 8 percentage points over random FedAvg in mid-scale configurations, and SqueezeNet benefited most when many clients were available but few could be selected per round. ResNet-18, with its deeper representational capacity, was largely indifferent to the selection strategy, showing negligible differences. EfficientNet behaved in the opposite direction, favoring random FedAvg in large federations of 50 to 100 clients, apparently because its strong generalization thrives on the diversity that random sampling introduces, while PSO’s targeted selection narrows the model’s exposure to rare but informative patterns. Pareto front analysis, which connects configurations that cannot be improved in accuracy without worsening execution time, showed that PSO contributed more frequently to the efficiency frontier in moderate-scale setups, reinforcing its value where computational and communication budgets are tight.

An ablation study dissected where the gains originate. Replacing random sampling with the complete FedPSO module raised mean final accuracy from 94.26 to 94.57 percent and macro F1 from 93.76 to 94.11 percent, beating the baseline in 18 of 28 paired configurations. Yet a greedy selector guided purely by empirical accuracy reached 94.78 percent, revealing that client utility is the dominant ingredient and that the swarm provides flexible subset optimization rather than a simple ranking. A Genetic Algorithm baseline exploiting the same utility information landed at 94.32 percent, confirming that metaheuristics broadly benefit from utility-guided selection while FedPSO stays simpler and fully FedAvg-compatible. A factorial grid search established a stable practical parameter interval, with the best SqueezeNet configuration reaching 98.89 percent accuracy at an inertia weight of 0.9 and the best MobileNet setting 97.25 percent, within an interval of 0.7 to 0.9 for the inertia weight, 10 to 20 particles and 500 to 1000 evaluations. A Dirichlet stress test exposed a boundary condition: under strong label skew with α = 0.3, FedPSO reached 98.55 percent on SqueezeNet in a 30-client federation, but performance degraded substantially when 100 clients offered only five participants per round, because sparse participation combined with severe skew drives the optimizer to over-exploit a narrow subset.

For clinical deployments, the implications are concrete. Small hospitals and rural clinics often contribute limited but highly specialized diagnostic data, and the ability to prioritize such institutions can sharpen model sensitivity for underrepresented conditions without any additional data collection or centralization. Because lightweight architectures such as SqueezeNet and MobileNet are the natural choice for edge devices and mobile diagnostic platforms, and because they benefit the most from intelligent selection, the strategy aligns well with federated medical AI systems operating under strict computational and communication constraints. The authors plan to extend the framework to more complex clinical tasks, to integrate additional metaheuristics and reinforcement learning, and to add diversity-aware fitness terms that periodically include lower-ranked clients, mitigating the optimizer’s tendency to converge on a narrow participant subset. The broader message is that in privacy-preserving medicine, who trains the model can matter as much as how the model is trained.

Subject of Research: Dynamic client selection in federated learning using particle swarm optimization for medical image classification

Article Title: Dynamic client selection in federated learning using particle swarm optimization for medical image classification

Article References: Rodrigues, L. G. F., Rodrigues Moreira, L. F., Moreira, R., & Backes, A. R. (2026). Dynamic client selection in federated learning using particle swarm optimization for medical image classification. Cluster Computing, 29(15), Article 885. https://doi.org/10.1007/s10586-026-06640-6

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06640-6

Keywords: federated learning, particle swarm optimization, medical imaging, client selection, deep learning, data privacy, convolutional neural networks, non-IID data, Dynamic, client, selection, federated