Siloed Federated Learning for Multi-Centric Histopathology Datasets

• Inter-center distribution differences can bias models or prevent convergence, and histopathology commonly varies in staining, scanner configuration, and imaging artifacts.Source: §1 · p. 2
• The paper demonstrates federated tile-level histopathology image classification with Camelyon16 and Camelyon17.The authors describe this as among the first such examples to their knowledge.Source: §1 · p. 2
• SiloBN uses local-statistic batch-normalization layers to address inter-center variability without communicating local activation statistics.The authors say this reveals less sensitive information.Source: §1 · p. 2
• SiloBN achieves the same or better intra-center generalization than existing federated techniques and better out-of-domain generalization.The introduction describes these results as holding even in challenging settings.Source: §1 · p. 2

• Federated learning optimizes a shared model over participant losses, using sampled participants and local optimization before model aggregation.FedAvg performs several local optimization steps before aggregation to reduce coordination rounds.Source: §2 · pp. 2–3
• FedAvg has no natural convergence guarantee on highly dissimilar participant datasets, while FedProx adds a quadratic loss term to restrain local-model divergence.Other approaches adapt participant sampling or cluster participants into effective collaborators.Source: §2 · p. 3
• The paper retains an i.i.d. assumption within each participant, but not across participants.Source: §2 · p. 3

• Medical collaborative federated learning involves few centers with sizable, higher-quality local datasets, unlike consumer federated learning with many participants and comparatively little data each.Source: §3 · p. 3
• In collaborative federated learning, participants seek to improve their local task performance through collaboration and keep sensitive evaluation data at their centers.Consumer federated learning instead centers on a service provider distributing a global model to participants.Source: §3 · pp. 3–4
• The study trains a tile-level whole-slide image classifier across centers, using full-participation rounds because strong participant subsampling can bias updates and raise privacy challenges.The authors use E to denote the number of local mini-batch updates per round.Source: §3 · p. 4

Batch Normalization

• BN renormalizes an input tensor using channel-wise running means and variances, alongside learned affine parameters.The means and variances are computed across spatial and batch dimensions.Source: §4 · p. 4

SiloBN

• Unlike naive FedAvg with BN layers, which aggregates both BN statistics and learned parameters, SiloBN separates their treatment.BN statistics encode local domain information, while learned BN parameters can transfer across domains.Source: §4 · pp. 4–5
• SiloBN shares learned BN parameters across centers while keeping BN statistics local and sharing non-BN parameters in the standard fashion.The local statistics help center-specific intermediate activations remain similarly centered across centers.Source: §4 · p. 5

Model Personalization and Transfer

• SiloBN learns one model per center, and AdaBN can adapt it to an unseen target by recomputing BN statistics while freezing other parameters.The recomputation uses a data batch from the new target domain.Source: §4 · p. 5

Privacy and BN

• Because SiloBN does not share BN statistics, the central server never accesses all network parameters, reducing shared information that could aid privacy attacks.The passage states that this can only help diminish the effectiveness of attacks that are most effective with full-parameter knowledge.Source: §4 · p. 5

5.1 Datasets

• FL-C16 and FL-C17 classify tumorous versus healthy tiles from Camelyon images drawn from two and five hospitals, respectively.Tiles are partitioned into training, validation, and test sets using per-hospital stratification; tiles from the same slide in FL-C16 or patient in FL-C17 stay in one partition.Source: §5.1 · p. 6

5.2 Experiments on FL Histopathology Datasets

• The study compares FedAvg, FedProx, pooled, and independent local training with SiloBN, using DCNNs with and without BN and held-out center-specific AUCs.Experiments vary local updates per federated round; mAUC is the pan-center mean of intra-center AUCs.Source: §5.2 · pp. 6–7
• SiloBN outperforms other federated strategies with DCNN+BN on both datasets and is on par with pooled and local training; without BN, FedAvg and SiloBN are equivalent.The full table reports mean AUC across centers and shows SiloBN as the most effective approach when using BN layers.Source: §5.2 · p. 8, Table 1 · p. 8

5.3 Out-of-domain generalization experiments

• From FL-C16 to FL-C17, SiloBN with AdaBN outperforms FedAvg-trained DCNN on average and is more stable across trainings and centers.The table summarizes domain generalization from FL-C16 to FL-C17.Source: §5.3 · p. 8, Table 2 · p. 8

Conclusion

• SiloBN yields similar or better intra-center generalization than existing FL methods on real-world multicentric histopathology datasets.Source: §6 · p. 9
• SiloBN outperforms existing methods on out-of-domain generalization in both performance and stability.Source: §6 · p. 9
• By not sharing local activation statistics, SiloBN has a better privacy profile; future work could quantify the privacy benefits and study other domains such as radiology.Source: §6 · p. 9

A Supplementary Material

• Table 3 gives the distribution of healthy and tumor-containing tiles across centers for each federated dataset.The caption provides no tile counts.Source: Table 3 · p. 11
• For each dataset, local training uses E = 1; the table also reports validation-selected hyperparameters, including BN momentum 0.1 when applicable and eps = 10^-5.The Adam first- and second-moment decay hyperparameters are denoted β1 and β2.Source: Table 4 · p. 11
• Figure 3 shows the DCNN architecture used in all experiments; convolutions use spatial zero-padding of size 1, and convolutional and linear layers include biases.Each convolution’s output-channel count corresponds to its block’s output-channel count.Source: Figure 3 · p. 12
• Figure 4 presents four representative WSIs for each identified center in the Camelyon16 and Camelyon17 datasets.Source: Figure 4 · p. 12