Research map / Federated learning and privacy / Personalized and heterogeneous FL
HistoFS: Non-IID Histopathologic Whole Slide Image Classification via Federated Style Transfer with RoI-Preserving
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HistoFS, a federated MIL framework for WSI classification that augments local bags with multiple clustered "pseudo bag styles" via AdaIN and uses an attention-based authenticity module to keep RoIs intact.
Problem
Non-i.i.d. feature shifts across institutions in federated multiple instance learning for whole slide image classification, where single-style cross-client style transfer ignores multiple WSI morphologies and can shift regions of interest.
Approach
Cluster each WSI's patch-level style distributions (mean/std) into K pseudo bag styles via 2-Wasserstein K-means, share them across clients for AdaIN augmentation at the bag level, and align attention weights between original and augmented bags with a cosine-similarity authenticity score to preserve RoIs.
Key results
- HistoFS achieves average AUC of 98.11% on RCC and 86.07% on HER2, the highest among compared methods (e.g., CCST 97.26% and 81.18%).
- Entropy score of AUC across institutions is lowest for HistoFS (0.056 RCC, 0.383 HER2), indicating reduced performance divergence.
- On unseen Camelyon17 clients, HistoFS reaches 84.67% average AUC versus 82.30% for CCST, with communication load about 197 MB versus 420 MB for CCST on RCC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add pseudo bag styles transfer and the authenticity module | Attention-gated MIL in federated WSI classification · RCC | AUC (%): 96.99 -> 98.11 | Helped |
| add pseudo bag styles transfer and the authenticity module | Attention-gated MIL in federated WSI classification · HER2 | AUC (%): 82.31 -> 86.07 | Helped |
| add pseudo bag styles transfer and the authenticity module | Attention-gated MIL in federated WSI classification · Camelyon17 unseen clients average | AUC: 82.30 -> 84.67 | Helped |
| add the authenticity module to pseudo bag styles transfer (ablation) | HistoFS · RCC | AUC (%): 97.81 -> 98.11 | Helped |
| add the authenticity module to pseudo bag styles transfer (ablation) | HistoFS · HER2 | AUC (%): 84.98 -> 86.07 | Helped |
| add the authenticity module to pseudo bag styles transfer (ablation) | HistoFS · Camelyon17 unseen clients average | AUC: 82.52 -> 84.67 | Helped |
| replace attention-gated aggregation with DTFD-MIL (ablation) | HistoFS · RCC | AUC (%): 98.11 -> 97.79 | Hurt |
| replace attention-gated aggregation with DTFD-MIL (ablation) | HistoFS · HER2 | AUC (%): 86.07 -> 85.33 | Hurt |
| replace attention-gated aggregation with FRMIL (ablation) | HistoFS · RCC | AUC (%): 98.11 -> 97.95 | Hurt |
| replace attention-gated aggregation with FRMIL (ablation) | HistoFS · HER2 | AUC (%): 86.07 -> 85.49 | Hurt |
| replace attention-gated aggregation with DTFD-MIL (ablation) | HistoFS · Camelyon17 unseen clients average | AUC: 84.67 -> 82.94 | Hurt |
| replace attention-gated aggregation with FRMIL (ablation) | HistoFS · Camelyon17 unseen clients average | AUC: 84.67 -> 83.89 | Hurt |
Surprising findings
- Local MixStyle and DSU and federated CCST and DACS give only minor gains over the HistoFL baseline on unseen clients.
- DACS performs worse than local MixStyle on the HER2 seen-client benchmark in average AUC.
Limitations
- Relies on frozen SSL-ViT patch features and an attention-based MIL aggregator; evaluated only with ResNet-free SSL-ViT extractor variants in ablations.
- Authors note that the HER2 dataset remains harder to achieve fairness on (higher entropy score).
- Privacy-preserving joint design is left for future work; sharing style statistics is not formally protected.
Experimental setup
Datasets: HER2 (HEROHE GC, Yale HER2, TCGA-BRCA) RCC (TCGA) Camelyon17
Benchmarks: HER2 RCC Camelyon17
Metrics: AUC balanced accuracy F1-score entropy score of per-institution AUC communication load training time
Baselines: HistoFL DTFD-MIL FRMIL MixStyle DSU CCST DACS FedAvg
Builds on: AdaIN ABMIL DTFD-MIL HistoFL CCST
Closest papers by meaning
- Nonparametric Distribution Matching for Self-Supervised Whole-Slide Image Condensation (NeurIPS 2026 · similarity 0.88)
- Dynamic Graph Representation with Knowledge-aware Attention for Histopathology Whole Slide Image Analysis (CVPR 2024 · similarity 0.88)
- PathVQ: Reforming Computational Pathology Foundation Model for Whole Slide Image Analysis via Vector Quantization (NeurIPS 2025 · similarity 0.87)
- Rethinking Transformer for Long Contextual Histopathology Whole Slide Image Analysis (NeurIPS 2024 · similarity 0.86)
- Morphological Prototyping for Unsupervised Slide Representation Learning in Computational Pathology (CVPR 2024 · similarity 0.86)
- Unsupervised Foundation Model-Agnostic Slide-Level Representation Learning (CVPR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.