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

Muhammad Muhtasim Shahriar; M. M. Golam Hafiz; Saad Aloteibi; Mohammad Ali Moni

cs.CVSep 25, 2026

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

Authors: Muhammad Muhtasim Shahriar, M. M. Golam Hafiz, Saad Aloteibi, Mohammad Ali Moni

Organizations: Department of Computer Science, International Islamic University Chittagong (IIUC), Chittagong, Bangladesh · Department of Computer Science, American International University-Bangladesh (AIUB), Dhaka, Bangladesh · Department of Computer Science and Engineering, College of Applied Studies, King Saud University, Riyadh, 11437, Saudi Arabia · AI & Digital Health Technology, Artificial Intelligence and Cyber Futures Institute, Charles Sturt University, Bathurst, NSW, 2795, Australia ·

Abstract

Cross-site lung histopathology classification must account for stain variation, non-IID client data, missing classes, and the cost of adapting large pathology encoders. This study evaluates FedHisto-PAST v2 for three-way classification of adenocarcinoma (ACA), Normal, and squamous cell carcinoma (SCC). FedHisto-PAST v2 combines a frozen HIBOU-B foundation model with parameter-efficient adaptation, stain-conditioned paired-view prediction and feature consistency, reliability-aware prototype learning, and adaptive federated aggregation. Experiments used a five-client, non-IID, raw-data-local simulation with fixed internal evaluation, client-level analysis, component ablations, communication accounting, and a development-influenced exploratory LungHist700 cohort. All principal methods achieved near- ceiling internal performance, which limited discrimination on the fixed split. On LungHist700, FedHisto- PAST v2 achieved a Macro-F1 of 0.728560 and a balanced accuracy of 0.730454. Higher recognition of Normal and SCC was accompanied by lower ACA recall, and calibration remained imperfect. Prediction-level consistency was the only component with a clearly supported independent contribution in the external ablation analysis. Feature consistency and prototype regularization showed no conclusive independent overall gains in Macro-F1. The framework updated 1.253841% of the model parameters. The results provide exploratory cross-dataset evidence for stain-aware, parameter-efficient federation; they do not establish formal privacy, patient-level independence, prospective deployment, or clinical validation.

Figures & tables

Study	Primary pathology task	Federated/non-IID strategy	Stain handling	Foundation model or PEFT	Prototype/consistency mechanism	Formal privacy mechanism evaluated?	External or unseen-domain evaluation	Remaining gap relative to FedHisto-PAST v2
SiloBN [ 5 ]	Multicentric tumour-patch classification	Centre-specific batch-normalization statistics with shared remaining parameters	Local normalization statistics	Not reported	Not reported	No formal mechanism evaluated	Cross-dataset transfer between Camelyon cohorts	No pathology foundation-model PEFT, paired stain consistency, or reliability-aware prototypes
Orchestral stain-normalization GAN [ 7 ]	Histopathology image analysis under distributed staining styles	Federated learning with orchestral generative stain normalization	GAN-based stain normalization	Not reported	No class-prototype mechanism reported	No formal mechanism evaluated	Multi-source evaluation reported	Focuses on normalization rather than paired feature/prediction consistency and missing-class-aware prototypes
BottleGAN [ 8 ]	Federated computational-pathology segmentation	Federated unsupervised stain mapping combined with weight aggregation	Many-one-many generative stain normalization	Not reported	No prototype or paired-consistency mechanism reported	No formal mechanism evaluated	Heterogeneous multi-institutional evaluation	Requires a generative normalization stage and does not address foundation-model PEFT or missing classes
FedSDA [ 9 ]	Non-IID histopathology classification	Federated stain-distribution alignment using stain matrices and diffusion modelling	Stain separation and distribution alignment	Not reported	No class-prototype mechanism reported	Privacy-risk mitigation discussed; no formal guarantee evaluated	Cross-client and non-IID evaluation reported	Distribution alignment is not paired counterfactual consistency and does not federate lightweight foundation-model adapters
FedWSIDD [ 35 ]	Weakly supervised whole-slide classification	Clients exchange distilled synthetic slides and may use heterogeneous MIL models	Stain normalization inside dataset distillation	Pretrained feature extraction; no federated PEFT reported	No class-prototype or paired-consistency mechanism reported	No formal mechanism evaluated	Camelyon16 and Camelyon17 multi-centre evaluation	Targets model heterogeneity through synthetic data exchange rather than stain-conditioned PEFT and reliability-aware semantic regularization
FedHisto-PAST v2	Three-class ACA/Normal/SCC image classification	Five-client non-IID simulation with adaptive reliability-aware aggregation	Client stain descriptors and paired stain-counterfactual views	Frozen HIBOU-B with lightweight trainable adaptation	Prediction/feature consistency and class-availability/reliability-aware prototypes	No formal mechanism; raw-data-local simulation	Fixed internal evaluation and development-influenced exploratory LungHist700 evaluation	Integrated target framework; prospective clinical validation and formal privacy guarantees remain outside scope

Table 1: Positioning of closely related federated computational-pathology methods and FedHisto-PAST v2

Partition/client	Pre-holdout ACA/Normal/SCC	Optimization ACA/Normal/SCC	Holdout ACA/Normal/SCC	Principal non-IID characteristic
C1	229/208/293	206/187/264	23/21/29	Relatively balanced; small
C2	485/439/65	437/395/59	48/44/6	SCC minority
C3	187/172/318	168/155/286	19/17/32	SCC enriched
C4	64/58/553	58/52/498	6/6/55	SCC majority
C5	1181/1831/0	1004/1556/0	177/275/0	External-source; SCC absent
Global validation	472/474/473	N/A	N/A	Fixed; approximately balanced

Table 2: Compact client and partition composition

Method	Adaptation/conditioning	Consistency	Prototype mechanism	Aggregation	Executed evidence
FedAvg-PEFT	LoRA + adapter + classifier	None	None	Sample-weighted	Internal 5; external 5
FedProx-PEFT	FedAvg capacity; proximal loss	None	None	Sample-weighted; $\mu=0.001$	Internal 4; external 5
FiLM-Zero	PEFT + zero-conditioned FiLM	None	None	Sample-weighted	Internal 5; external 5
FedHisto-PAST v1	PEFT + stain FiLM	Single-view	Availability-aware v1	Adaptive	Internal 5; external whole-image 5; patch-grid sensitivity secondary
FedHisto-PAST v2	PEFT + stain FiLM	Paired JS + feature	Reliability-aware cosine-margin	Adaptive	Internal 5; external whole-image 5
No prediction consistency	Full method otherwise	JS removed; feature retained	Unchanged	Unchanged	Internal 3; external 3 (seeds 11/33/55)

Table 3: Compact comparator and ablation design

Figure 1: Integrated FedHisto-PAST study and learning framework. (A) Canonical data clearance, C1–C5 non-IID partitions, fixed global validation and test sets, client holdouts, and the internal, ablation, and exploratory evidence tiers. (B) Shared-geometry stain-counterfactual views processed by frozen HIBOU-B, LoRA, residual adaptation, stain-conditioned FiLM, and supervised, prediction-consistency, feature-consistency, and prototype losses. (C) Local optimization, trainable-state transfer, class-availability masks, reliability-aware prototype exchange, adaptive model aggregation, and delta-SVD LoRA refactorization. (D) Validation-based checkpoint selection, fixed internal and client-level testing, five-checkpoint LungHist700 ensembling, and image-level paired bootstrap analysis without patient-clustered inference.

Method	Finalized $n$	Macro-F1	Accuracy	Balanced accuracy	ECE	Brier	SCC recall	Worst-client Macro-F1	Execution status
FedAvg-PEFT	5	$0.999012\pm 0.001070$	$0.999014\pm 0.001068$	$0.999012\pm 0.001070$	$0.001084\pm 0.000863$	$0.001817\pm 0.001790$	$0.998305\pm 0.002321$	$0.988713\pm 0.006396$	Complete
FedProx-PEFT	5	$0.998589\pm 0.001577$	$0.998592\pm 0.001575$	$0.998588\pm 0.001579$	$0.000999\pm 0.000816$	$0.001851\pm 0.001506$	$0.996610\pm 0.004130$	$0.985585\pm 0.010183$	Complete
FiLM-Zero	5	$0.998449\pm 0.001684$	$0.998451\pm 0.001681$	$0.998448\pm 0.001685$	$0.001316\pm 0.001244$	$0.002298\pm 0.002364$	$0.997034\pm 0.003545$	$0.979749\pm 0.012749$	Complete
FedHisto-PAST v1	5	$0.998872\pm 0.000631$	$0.998873\pm 0.000630$	$0.998872\pm 0.000630$	$0.001055\pm 0.000560$	$0.001836\pm 0.001164$	$0.998729\pm 0.001160$	$0.994357\pm 0.007763$	Complete
FedHisto-PAST v2	5	$0.997321\pm 0.002140$	$0.997324\pm 0.002136$	$0.997318\pm 0.002141$	$0.002627\pm 0.002075$	$0.004843\pm 0.003653$	$0.994915\pm 0.005318$	$0.977155\pm 0.019807$	Complete

Table 4: Execution coverage and internal fixed-split evaluation summary

Method	Accuracy	Balanced accuracy	Macro-F1	Macro-AUROC	Macro-AUPRC	NLL	Brier	ECE	ACA recall	Normal recall	SCC recall
FedAvg-PEFT	0.677279	0.678285	0.674942	0.902991	0.841937	4.074109	0.590220	0.273546	0.989286	0.741722	0.303846
FedProx-PEFT	0.693198	0.691828	0.696334	0.923432	0.886197	2.173750	0.541572	0.251508	0.975000	0.735099	0.365385
FiLM-Zero	0.674385	0.672019	0.672025	0.901589	0.847102	3.610070	0.593431	0.282894	0.989286	0.715232	0.311538
FedHisto-PAST v1	0.677279	0.681427	0.677191	0.875529	0.806098	5.487567	0.587681	0.287265	0.975000	0.761589	0.307692
FedHisto-PAST v2	0.717800	0.730454	0.728560	0.913336	0.848381	3.648615	0.513336	0.248052	0.932143	0.847682	0.411538

Table 5: LungHist700 whole-image probability-ensemble performance

Figure 2: Exploratory LungHist700 whole-image ensemble performance. (A) Macro-F1 and balanced accuracy across the five evaluated federated configurations. (B) Class-wise recall for adenocarcinoma (ACA), Normal, and squamous cell carcinoma (SCC). FedHisto-PAST v2 improved Normal and SCC recall while reducing ACA recall relative to the whole-image comparators.

Comparator	Macro-F1 difference 95% CI	Two-sided bootstrap $p$	Recall difference SCC / ACA
FedAvg-PEFT	0.053618 [0.027328, 0.079784]	$p<0.0002$	0.107692 $-0.057143$
FedProx-PEFT	0.032226 [0.008432, 0.056709]	$p=0.0090$	0.046154 $-0.042857$
FiLM-Zero	0.056535 [0.029800, 0.083942]	$p<0.0002$	0.100000 $-0.057143$
FedHisto-PAST v1 (secondary patch-grid pipeline)	0.189335 [0.149462, 0.231351]	$p<0.0002$	0.184615 $-0.039286$

Table 6: Paired image-level LungHist700 bootstrap comparisons

Configuration	Internal Macro-F1	External Macro-F1	Balanced accuracy	ACA recall	Normal recall	SCC recall	Full v2 $-$ ablation	95% CI	Bootstrap $P(\Delta>0)$
Full v2 (seeds 11/33/55)	$0.997885\pm 0.002444$	0.726936	0.730281	0.925000	0.854305	0.411538	Reference	—	—
Without prediction consistency	$0.996944\pm 0.001077$	0.703363	0.714530	0.939286	0.854305	0.350000	0.023573	[0.009206, 0.039099]	1.0000
Without feature consistency	$0.997886\pm 0.001867$	0.732073	0.734484	0.925000	0.847682	0.430769	$-0.005138$	[ $-0.017560$ , 0.006767]	0.2072
Without prototype regularization	$0.997415\pm 0.001469$	0.737960	0.736469	0.939286	0.827815	0.442308	$-0.011024$	[ $-0.025819$ , 0.002940]	0.0614

Table 7: Three-seed internal Macro-F1 and external ensemble component-ablation results

Figure 3: Error and component-ablation analysis on LungHist700. (A) FedHisto-PAST v2 confusion matrix, with rows denoting true classes and columns denoting predicted classes. The dominant error was SCC misclassification as ACA. (B) Paired external Macro-F1 differences between full FedHisto-PAST v2 and each component ablation. Points denote full-v2 minus ablation differences, and horizontal lines show 95% image-level bootstrap confidence intervals from 5,000 replicates.

Method	Total parameters	Trainable parameters	Trainable fraction (%)	Trainable payload (MiB)	Communication per round (MiB)	Overhead vs FedAvg-PEFT (%)
FedAvg-PEFT	86,829,763	987,907	1.137752	3.768566	37.685661	0.000000
FedProx-PEFT	86,829,763	987,907	1.137752	3.768566	37.685661	0.000000
FiLM-Zero	86,829,763	1,088,707	1.253841	4.153088	41.530876	10.203390
FedHisto-PAST v1	86,829,763	1,088,707	1.253841	4.153088	41.615891	10.428978
FedHisto-PAST v2	86,829,763	1,088,707	1.253841	4.153088	41.615891	10.428978

Table 8: Parameter and communication efficiency

Histopathology Image Classification

Histopathology image classification uses computational models to identify tissue types, tumors, grades, and biomarkers from stained pathology images.

See all 146 papers

FedStain: Modeling Higher-Order Stain Statistics for Federated Domain Generalization in Computational Pathology
May 14, 2026Fengyi Zhang, Junya Zhang, Wenzhuo SunComputational Pathology Federated Domain Generalization
School of Electronic Science and Technology, Hainan University, Haikou, China, 570228 · School of Computer Science and Technology, Xidian University, Xi’an, China, 710126 · Xiangjiang College of Elite Engineers, Hunan University, Changsha, China, 410082
Federated Distillation for Whole Slide Image via Gaussian-Mixture Feature Alignment and Curriculum Integration
May 1, 2026Luru Jing, Cong Cong, Yanyuan Chen +1Computational Pathology Dataset Distillation
School of Computer Science, Peking University, Beijing, China · Center for Health Informatics, Australian Institute of Health Innovation, Macquarie University, Sydney, NSW 2113, Australia · School of Data Science, University of Virginia, Charlottesville, VA, USA.
Decentralized Gossip Learning and Federated Averaging for Histopathology Image Classification
Sep 15, 2026Yusuf Ozturk, Bengisu Atli, Enes Goktekin +2Federated Learning Aggregation Non-IID Federated Learning
Antalya Bilim University, Electrical and Electronics Engineering Dep. Antalya, Türkiye · Antalya Bilim University, Computer Engineering Dep. Antalya, Türkiye · Ankara University, Graduate School of Natural and Applied Sciences, Ankara, Türkiye +1

May 14, 2026J/K move · Enter open · S save

FedStain: Modeling Higher-Order Stain Statistics for Federated Domain Generalization in Computational Pathology

Fengyi Zhang, Junya Zhang, Wenzhuo Sun

School of Electronic Science and Technology, Hainan University, Haikou, China, 570228 · School of Computer Science and Technology, Xidian University, Xi’an, China, 710126 · Xiangjiang College of Elite Engineers, Hunan University, Changsha, China, 410082

Computational Pathology Federated Domain Generalization Whole-Slide Image Analysis Federated Learning

Open paper AskPDF arXiv

Robust whole-slide image (WSI) analysis under strict data-governance remains challenging due to substantial cross-institutional stain heterogeneity. Domain generalization (DG) mitigates these shifts but typically requires centralized data, conflicting with privacy regulations. Federated learning (FedL) provides a decentralized alternative; however, existing FedL and federated DG (FedDG) approaches rely almost exclusively on low-order statistics, assuming Gaussian-like stain distributions. In contrast, real-world staining processes often produce asymmetric, heavy-tailed color distributions due to biochemical diffusion and scanner nonlinearity. Consequently, current methods fail to model the higher-order, non-Gaussian characteristics dominating real-world stain variability. To address this, we propose FedStain, a stain-aware FedDG framework explicitly incorporating higher-order stain moments--skewness and kurtosis--as compact statistical descriptors exchanged during federated optimization. These descriptors require no pixel-level data transmission, preserving strict privacy and communication efficiency, while enabling the global model to capture stain variability missed by low-order statistics. FedStain also employs a contrastive, cross-site parameter aggregation strategy to promote stain-invariant representations without relaxing data constraints. Extensive experiments on Camelyon17 and our new MvMidog-Fed benchmark show FedStain yields consistent improvements, outperforming state-of-the-art FedL, DG, and FedDG baselines by up to +3.9% absolute accuracy. To our knowledge, FedStain is the first FedDG approach to explicitly model higher-order stain statistics, enabling robust cross-institutional deployment in computational pathology.

Abstract

Histopathology Image Classification

Explore similar work

FedStain: Modeling Higher-Order Stain Statistics for Federated Domain Generalization in Computational Pathology

Federated Distillation for Whole Slide Image via Gaussian-Mixture Feature Alignment and Curriculum Integration

Decentralized Gossip Learning and Federated Averaging for Histopathology Image Classification