FedHisto-PAST: Parameter-Efficient Stain-Aware Federated Learning for Cross-Site Lung Histopathology Classification

Problem and motivation

• Cross-site classification must remain robust across laboratories, staining workflows, scanners, and under-represented classes, beyond high accuracy on curated images.The framework responds to a gap in jointly addressing stain shift, label skew, and adaptation cost.Source: §1 · pp. 1–2
• Existing approaches address individual aspects, but their interaction in a raw-data-local lung histopathology framework remains unresolved.Prior work includes stain normalization and stain-distribution alignment, while foundation-model features may retain site signatures and large-model adaptation increases trainable state.Source: §1 · pp. 1–2

Proposed framework

• FedHisto-PAST v2 uses a frozen HIBOU-B encoder with lightweight adaptation, paired stain-counterfactual views, and class-availability- and reliability-aware prototypes.It combines prediction and representation consistency with prototype regularization while retaining a shared global decision space when a class is absent locally.Source: §1 · p. 2

Contributions and evaluation

• The study contributes an integrated architecture and evaluates it through multi-seed internal testing, client robustness, calibration, exploratory cross-dataset analysis, and communication accounting.Component-specific claims are restricted to the reported ablation evidence, and the study distinguishes raw-data locality from formal privacy.Source: §1 · p. 2

2.1. Computational Lung Histopathology Classification and Stain Variation

• Lung histopathology models distinguish tissue categories, but curated-source performance does not establish robustness to laboratory shift or other site-linked appearance cues.Cross-site methods must separate disease morphology from stain-dependent appearance while preserving minority-class evidence.Source: §2.1 · pp. 2–3

2.2. Federated Learning in Computational Pathology

• Federated methods address local optimization, normalization, and pathology tasks, while stain-aware approaches use translation or distribution alignment rather than paired counterfactual views [10] [29] [30].Raw-data locality alone does not provide differential privacy or secure aggregation.Source: §2.2 · p. 3

2.3. Pathology Foundation Models and Parameter-Efficient Adaptation

• Pathology foundation models offer transferable representations, but site-specific signals can impair unseen-domain performance; low-rank adaptation reduces trainable state without addressing heterogeneity by itself [11].Federated adaptation therefore needs to connect lightweight updates with stain-aware objectives and meaningful aggregation under missing classes.Source: §2.3 · p. 3

2.4. Consistency Learning, Prototype Methods and Reliability Evaluation

• Paired stain-counterfactual consistency compares views of the same tissue, while prototypes transfer class-level summaries whose reliability depends on local evidence [10].Evaluation should also report client-level behavior, calibration, communication cost, and external-domain performance.Source: §2.4 · pp. 3–4

2.5. Synthesis and Research Gap

• The identified gap is a unified lung-histopathology framework combining parameter-efficient foundation-model adaptation, stain conditioning, paired-view consistency, reliable prototypes, adaptive aggregation, and auditable evaluation.FedHisto-PAST v2 studies this combination in a five-client non-IID simulation with exploratory cross-dataset evaluation and no formal privacy guarantee.Source: §2.5 · p. 4

3.1. Study Design and Reporting Scope

• The study uses five simulated clients to classify ACA, Normal, and SCC, keeping source images and labels within assigned partitions while coordinating a frozen encoder and lightweight modules.The protocol is a dataset-partition simulation, not a study across operational hospitals.Source: §3.1 · p. 4
• Fixed global validation and test partitions support checkpoint selection and final evaluation, while LungHist700 is treated as development-influenced exploratory evidence.The protocol is neither leave-one-client-out nor cross-validation, and formal privacy mechanisms were outside its scope.Source: §3.1 · p. 4

3.2. Dataset Construction and Client Partitioning

• The canonical corpus contained 8,922 labelled images, partitioned into client pools and fixed global validation and test sets; client distributions included a C5 partition with no SCC images.C1–C4 used 10% local holdouts, compared with 15% for C5.Source: §3.2 · p. 5

3.3. Dataset Clearance, Leakage Control, and Split Integrity

• Path, exact-file, and similarity audits found no detected cross-boundary family overlaps under the implemented procedures, but unavailable patient and slide identifiers prevent claims of biological or clinical independence.The authors note that transformed copies may evade hashes and similarity procedures depend on image characteristics and thresholds.Source: §3.3 · p. 6

3.4. Image Preprocessing and Stain-Counterfactual View Generation

• A shared random geometric transformation creates matched views so consistency compares different stain realizations of the same tissue content.Evaluation uses deterministic whole-image resizing and normalization; the final method uses the paired-stain pathway.Source: §3.4 · p. 6
• The paired views receive independent optical-density and brightness/contrast perturbations designed to alter stain expression while preserving morphology.These perturbations were not treated as pathologist-validated stain normalization.Source: §3.4 · p. 6

3.5. Stain Descriptor and Stain-Conditioned FiLM

• A 14-dimensional client stain descriptor conditions a shared FiLM network, whose modulation scale is limited to 0.10.Client holdouts use their corresponding client descriptor; global and LungHist700 inference use a centroid computed from the five training-client descriptors.Source: §3.5 · pp. 6–7

3.6. Model Architecture and Parameter-Efficient Adaptation

• The architecture freezes pretrained HIBOU-B and adapts it with LoRA, a residual adapter, a classifier, and optional FiLM modules.LoRA is attached to attention query, key, value, and output projections; the paired branches share all trainable modules.Source: §3.6 · p. 7
• Both stain-counterfactual views pass through the same encoder and adaptation pathway, with supervised classification combined with prediction, feature, and prototype regularization.Class-balanced cross-entropy is applied symmetrically; a class absent locally receives zero weight, with no replacement oversampling.Source: §3.6 · p. 7

3.7. Consistency Losses and Reliability-Aware Prototype Regularization

• Prediction consistency minimizes Jensen-Shannon divergence between view distributions, while feature consistency minimizes cosine distance between normalized adapted embeddings.Both views receive supervised losses; neither acts as a fixed teacher.Source: §3.7 · p. 7
• Clients compute class centroids only for locally observed classes, and the server aggregates prototypes only where class-availability masks are true.Prototype reliability uses server-side variance when available and a count-based fallback otherwise; samples without a valid global prototype are omitted.Source: §3.7 · p. 7
• The prototype regularizer combines reliability-weighted cosine attraction to the correct prototype with a hardest-available-negative margin of 0.20.Component weights and the margin were fixed before exploratory evaluation; removal experiments tested individual terms within the frozen configuration.Source: §3.7 · pp. 7–8

3.8. Federated Optimization, Aggregation, and Comparators

• Each communication round distributes the current trainable model state and applicable global prototypes; participating clients perform two local epochs and return updated trainable state and available-class prototypes.Canonical complete rounds retain all five planned clients, and the server aggregates returned states rather than fabricating missing updates.Source: §3.8 · p. 8
• FedAvg-PEFT uses sample-mass-weighted aggregation, while FedProx-PEFT adds a proximal penalty with coefficient μ = 0.001.FiLM-Zero uses a constant zero descriptor and no prototype mechanism; FedHisto-PAST v1 lacks paired-view consistency.Source: §3.8 · p. 8
• FedHisto-PAST's adaptive server rule weights updates using sublinear sample mass, agreement with a leave-one-out reference, and deviation from the median update norm.After normalization, a minimum client weight of 0.03 is enforced; the server combines reconstructed LoRA updates and refactorizes them by delta-SVD.Source: §3.8 · p. 8
• The adaptive aggregation rule is an empirical heuristic, not a trust score, Byzantine-robustness mechanism, or method with a new convergence proof.Model aggregation uses returned-update statistics, whereas prototype aggregation uses class availability, contribution counts, and reliability evidence.Source: §3.8 · p. 8

3.9. Training Control, Checkpointing, and Reproducibility

• Training allowed up to 50 communication rounds, with early stopping based on global-validation Macro-F1 and separate restoration of the best trainable and prototype states.Stopping required at least 10^-4 improvement and followed 10 nonimproving rounds.Source: §3.9 · p. 8
• Seeded execution, pinned model revision, canonical data manifest, configuration checks, atomic checkpoints, and prediction exports support reproducibility, but exact bitwise GPU replay was not guaranteed.Communication totals include trainable-state downloads and uploads across five clients, with auxiliary prototype traffic included where used; the one-time descriptor exchange is tracked separately.Source: §3.9 · p. 8

3.10. Internal Evaluation Protocol

• Each finalized seed's validation-selected checkpoint was evaluated once on the fixed global test set and client-specific holdouts, with Macro-F1 as the primary endpoint.Reporting also covered classification, discrimination, and calibration metrics; absent client classes remain an interpretation boundary.Source: §3.10 · pp. 9–10
• FedAvg-PEFT, FiLM-Zero, FedHisto-PAST v1, and v2 each had five finalized internal reports; FedProx-PEFT had four, with seed 44 not reconstructed or imputed.The seed-44 checkpoint remained eligible for frozen external inference; component ablations used seeds 11, 33, and 55.Source: §3.9 · p. 9, §3.10 · p. 10

3.11. LungHist700 Exploratory Cross-Dataset Evaluation and Statistical Analysis

• The exploratory LungHist700 manifest contains 691 images—280 ACA, 151 Normal, and 260 SCC—and uses deterministic whole-image resizing and normalization.No cohort image was used for optimization, aggregation, checkpoint selection, threshold adjustment, calibration, or post-freeze parameter modification.Source: §3.11 · p. 10
• Each method-level LungHist700 ensemble averages image-aligned probability vectors from five frozen checkpoints before assigning the argmax class.The same deterministic whole-image protocol was used for eligible methods in the primary comparison.Source: §3.11 · p. 10
• Paired comparisons resampled the same 691 images 10,000 times, but image-level resampling cannot establish patient-level uncertainty because patient identifiers were unavailable.The pairwise analyses were exploratory and used unadjusted probabilities without multiplicity correction.Source: §3.11 · p. 10

3.12. Privacy, Reproducibility, and Claim Boundaries

• Images and labels stayed within simulated-client partitions, but transmitted model states and summaries were not evaluated against update-level attacks and provide no formal privacy guarantee.The study did not implement differential privacy, secure aggregation, or other listed formal privacy mechanisms.Source: §3.12 · p. 10
• Claims are limited to the executed five-client image-level simulation, not patient- or slide-level independence, hospital operations, prospective deployment, diagnostic equivalence, or clinical readiness.The exploratory LungHist700 analysis is not presented as independent clinical validation.Source: §3.12 · p. 10

4.1. Execution Coverage and Evidence Integrity

• Across 25 planned principal runs, 24 internal reports were finalized; FedProx-PEFT seed 44 was excluded without score imputation, though its checkpoint remained available for frozen external inference.The failure occurred during checkpoint validation and reporting.Source: §4.1 · pp. 10–11

4.2. Internal Fixed-Split Performance

• All principal methods were near ceiling internally; v2 Macro-F1 was 0.997321 ± 0.002140, and rank differences did not establish its superiority.The fixed global test set was shared across methods and seeds.Source: §4.2 · p. 11, §4.1 · p. 11, Table 4 · p. 11

4.3. Client-Level Robustness

• V2 worst-client Macro-F1 was 0.977155 ± 0.019807, below v1's 0.994357 ± 0.007763, and ranged from 0.946208 to 0.998584 across seeds.Global-test performance and worst-client performance were retained as distinct evidence dimensions.Source: §4.3 · pp. 11–12

4.4. Exploratory LungHist700 Cross-Dataset Performance

• Macro-F1 0.728560 and balanced accuracy 0.730454: v2 exceeded whole-image comparators on Macro-F1, with higher Normal and SCC recall but lower ACA recall.The whole-image comparator Macro-F1 values were 0.674942 for FedAvg-PEFT, 0.696334 for FedProx-PEFT, 0.672025 for FiLM-Zero, and 0.677191 for v1.Source: Table 5 · p. 12, §4.4 · p. 12

4.5. Paired Image-Level Bootstrap Comparisons

• Paired image-level comparisons supported higher v2 Macro-F1 and SCC sensitivity than the three primary whole-image comparators, while ACA recall was lower; analyses were exploratory and unadjusted.The paired intervals supported those directions, but p-values were reported descriptively rather than as confirmatory evidence.Source: §4.5 · pp. 12–13, Table 6 · p. 13

4.6. Component Ablation Analysis

• Only removing prediction-level consistency produced an external Macro-F1 interval excluding zero in favor of full v2; feature consistency and prototype regularization showed no conclusive overall gains.Ablations used seeds 11, 33, and 55, with three-checkpoint external ensembles.Source: §4.6 · p. 13, Table 7 · p. 14

4.8. Calibration, Subgroup Behaviour and Error Patterns

• External ECE was 0.248052 and the probability scale remained imperfect; the dominant error was SCC misclassification as ACA (153 of 260), with no patient-clustered subgroup analysis.Subgroup summaries were descriptive and did not identify a biological cause for observed differences.Source: Figure 3 · p. 14, §4.8 · pp. 14–15

4.7. Parameter and Communication Efficiency

• V2 updated 1.253841% of parameters and communicated 41.615891 MiB per round across five clients, 10.428978% above FedAvg-PEFT; runtime and energy were unmeasured.The one-time stain-descriptor upload was approximately 0.000267 MiB.Source: §4.7 · p. 14, Table 8 · p. 15

5.1. Principal Findings

• Near-ceiling internal results did not establish superiority: FedHisto-PAST v2 did not lead the ranking, and its worst-client summary was lower and more variable than several comparators.The fixed split had limited resolution for distinguishing cross-site performance mechanisms.Source: §5.1 · p. 15
• On LungHist700, FedHisto-PAST v2 showed a different class-balance profile, with improved Normal and SCC recall alongside reduced ACA recall; bootstrap comparisons supported positive Macro-F1 and SCC-recall differences against three comparators.The comparisons were exploratory, unadjusted for multiplicity, and image-level; most errors were SCC images classified as ACA, without a supported biological or acquisition-specific explanation.Source: §5.1 · p. 15
• Prediction-level consistency had the clearest independently supported external contribution; feature consistency and prototype regularization showed no conclusive independent overall Macro-F1 gains.External calibration remained imperfect despite improved class balance.Source: §5.1 · p. 15

5.2. Interpretation in Relation to Prior Work

• Unlike stain normalization or distribution alignment, FedHisto-PAST v2 compares paired stain realizations of the same tissue; its ablation supports prediction consistency only within the tested configuration.The comparison does not establish that paired consistency is universally preferable across different tasks, cohorts, or evaluation units.Source: §5.2 · p. 16

5.3. Implications for Federated Computational Pathology

• Federated evaluation should pair fixed global testing with client-level summaries, class-specific sensitivity, calibration, and a changed source distribution, while considering class-specific error costs.Higher exploratory Normal and SCC recognition coincided with lower ACA recall.Source: §5.3 · p. 16
• Raw-data locality is not formal privacy protection, and the exploratory LungHist700 findings do not establish patient-level independence or prospective clinical validity.Keeping images within simulated client partitions may reduce direct data movement but is not equivalent to differential privacy, secure aggregation, or resistance to inference attacks.Source: §5.3 · p. 16

5.4. Limitations

• The study used five simulated partitions, lacked patient and whole-slide identifiers, and treated image-level separation audits as quality control rather than proof of patient-level independence.LungHist700 was development-influenced and exploratory; its image-level bootstrap may understate uncertainty when images are dependent within patients or slides.Source: §5.4 · p. 16
• The internal benchmark offered limited discrimination, and available FedHisto-PAST v1 patch-grid evidence could not isolate method differences from inference-pipeline differences.FedProx-PEFT had four finalized internal seed reports after a documented seed-44 report and checkpoint-validation failure; no value was reconstructed or imputed.Source: §5.4 · p. 16

6. Conclusion

• The fixed internal test set produced near-ceiling results for principal configurations, while exploratory LungHist700 results showed improved Macro-F1, balanced class recognition, Normal recall, and SCC sensitivity against available federated parameter-efficient comparators.The LungHist700 gains were accompanied by reduced ACA recall and imperfectly calibrated probabilities.Source: §6 · p. 17
• Prediction-level consistency was the only component with a clearly supported independent external contribution; feature consistency and prototype regularization had no conclusive independent overall Macro-F1 gains.The latter components remained part of the integrated formulation and affected class-specific behaviour.Source: §6 · p. 17
• FedHisto-PAST v2 updated approximately 1.25% of model parameters with a frozen foundation-model backbone, incurring modest but measurable communication overhead relative to FedAvg-PEFT.The supported conclusion is exploratory cross-dataset class-balance and SCC-sensitivity improvement under the tested image-level protocol.Source: §6 · p. 17
• The results do not establish universal superiority, clinical readiness, patient-level independence, or formal privacy; stronger claims require patient-grouped, multi-institutional prospective validation, calibration, and formal privacy evaluation where intended.Source: §6 · p. 17

• Muhammad Muhtasim Shahriar is credited with conceptualization, methodology, software, data curation, investigation, formal analysis, validation, visualization, and original-draft writing.Source: CRediT authorship contribution statement · p. 17
• M. M. Golam Hafiz is credited with data curation, investigation, formal analysis, validation, visualization, and original-draft writing.Source: CRediT authorship contribution statement · p. 17
• Saad Aloteibi is credited with supervision, project administration, methodology, resources, funding acquisition, and review-and-editing writing.Source: CRediT authorship contribution statement · p. 17
• Mohammad Ali Moni is credited with supervision, project administration, methodology, resources, and review-and-editing writing.Source: CRediT authorship contribution statement · p. 17

• The internal corpus used publicly available LC25000 and WSSS4LUAD images, while LungHist700 supported exploratory cross-dataset evaluation; no new clinical data were collected.The datasets are available through repositories associated with their original publications.Source: Data Availability · p. 17