Source-linked AI summary
Federated Learning for Computational Pathology on Gigapixel Whole Slide Images
Ming Y. Lu, Dehan Kong, Jana Lipkova, Richard J. Chen, Rajendra Singh, Drew F. K. Williamson, Tiffany Y. Chen, Faisal Mahmood
TL;DR
Computational pathology needs diverse, well-annotated data, but centralized integration of gigapixel whole slide images is constrained by privacy and data-sharing challenges. The paper combines federated learning with weakly supervised attention multiple instance learning and randomized weight noise, finding accurate classification and survival models can be trained across distributed institutions without direct data sharing.
Problem
Robust computational pathology models need diverse annotated data, while integrating multi-institutional whole slide images is limited by privacy concerns and data-sharing complexities.
Method
The study uses master-server federated learning with weakly supervised multiple instance learning, local model training, weight averaging, randomized noise, and slide- or patient-level labels.
Results
Federated learning produced accurate weakly supervised classification and survival models across distributed whole slide image datasets without direct patient-data sharing.
Takeaways & Limitations
The framework enables institutions to integrate whole slide image datasets while retaining local data control and applying randomized differential-privacy protection.
Takeaways & Limitations
The privacy guarantee depends on an assumption that Δ2(f) = 1 when relating noise level σ to the differential-privacy bound δ.
Abstract
from arXiv · showhide
Deep Learning-based computational pathology algorithms have demonstrated profound ability to excel in a wide array of tasks that range from characterization of well known morphological phenotypes to predicting non-human-identifiable features from histology such as molecular alterations. However, the development of robust, adaptable, and accurate deep learning-based models often rely on the collection and time-costly curation large high-quality annotated training data that should ideally come from diverse sources and patient populations to cater for the heterogeneity that exists in such datasets. Multi-centric and collaborative integration of medical data across multiple institutions can naturally help overcome this challenge and boost the model performance but is limited by privacy concerns amongst other difficulties that may arise in the complex data sharing process as models scale towards using hundreds of thousands of gigapixel whole slide images. In this paper, we introduce privacy-preserving federated learning for gigapixel whole slide images in computational pathology using weakly-supervised attention multiple instance learning and differential privacy. We evaluated our approach on two different diagnostic problems using thousands of histology whole slide images with only slide-level labels. Additionally, we present a weakly-supervised learning framework for survival prediction and patient stratification from whole slide images and demonstrate its effectiveness in a federated setting. Our results show that using federated learning, we can effectively develop accurate weakly supervised deep learning models from distributed data silos without direct data sharing and its associated complexities, while also preserving differential privacy using randomized noise generation.
1. Introduction
Computational pathology benefits from deep learning but requires diverse, well-annotated data, while whole-slide heterogeneity and privacy constraints complicate centralized multi-institutional training. The paper presents privacy-preserving federated learning combined with weakly supervised multiple instance learning for classification and survival prediction using distributed whole slide images.
- Deep learning supports pathology tasks spanning phenotype characterization, tumor-microenvironment quantification, survival prediction, and molecular alteration prediction.
- Whole slide images vary across patient populations, tissue preparation, staining protocols, and scanner hardware, motivating multi-institutional data integration.
- Centralizing medical data creates regulatory, legal, and privacy challenges, while model parameters or gradients can still expose sensitive information if leaked or attacked.
- Federated learning keeps sensitive patient data within institutional firewalls while transferring model specifics for local training and validation.
- The study demonstrates federated weakly supervised multiple instance learning for binary and multi-class classification using only slide-level labels.
- The framework also supports weakly supervised survival prediction from whole slide images using patient-level prognostic information without manual region-of-interest selection.
2. Methods
The method combines weakly supervised multiple instance learning with federated optimization and randomized noise to train privacy-preserving models on distributed gigapixel WSIs. WSIs are converted into patch-level embeddings, aggregated with attention into bag representations, and used for classification or discretized survival prediction.
- Federated learning: Federated training keeps institutional data local while periodically averaging client model parameters at a central server and returning synchronized weights.The framework uses a master-server architecture, with local training and parameter exchange after each epoch.
- Privacy protection: Randomized Gaussian noise is added to shared parameters to provide differential privacy, with the adjustable scale α linked to each layer’s parameter standard deviation.The privacy analysis assumes Δ_2(f)=1 and states that increasing σ improves the bound for δ.
- Data preprocessing: Each WSI becomes a bag of M 256 × 256 tissue crops, whose pretrained ResNet50 embeddings reduce the computational burden of gigapixel-scale analysis.Tissue regions are automatically segmented before cropping, and M varies with tissue content.
- Weakly-supervised learning: Attention MIL projects patch embeddings, scores their importance, and computes a weighted bag representation for downstream prediction.The model contains projection, attention, and prediction modules; attention weights are used to average feature embeddings.
- Survival prediction: Survival prediction discretizes event times into four quartile-based intervals and models conditional hazards from the patient or WSI bag representation.For patients with multiple WSIs, the individual bags are concatenated into a single patient bag before prediction.
- Federated deployment: Classification and survival models are trained locally within the federated setup, with client weights aggregated each epoch and protected using the Gaussian mechanism.The same federated procedure supports both predictive tasks while applying randomized protection to client-specific information.
3. Experiments and Results
The experiments evaluate privacy-preserving federated learning with weakly supervised whole-slide-image models across classification and survival prediction tasks. Results indicate strong multi-institutional performance, interpretability, and patient stratification, while excessive privacy noise can reduce classification performance.
- Experimental setup: WSIs are processed at 20× magnification as bags of smaller regions, allowing models to learn from slide- or patient-level labels without labeled regions of interest.The implementation uses patching and pretrained CNN feature extraction before local model training and federated aggregation.
- Dataset description: The classification datasets comprise 2126 breast WSIs and 1184 kidney WSIs, with institutional sites simulated by grouping tissue-source sites and in-house data.Data are partitioned by patient cases into training, validation, and test sets, with stratified sampling used to represent minority classes.
- Experiments: The study evaluates weakly supervised federated learning for BRCA and RCC classification and CCRCC survival prediction using distributed whole-slide-image data.The framework uses attention-based multiple instance learning and is evaluated across binary, multi-class, and survival tasks.
- Interpretability and survival results: Privacy-preserving federated models localized diagnostically relevant morphology in attention heatmaps and stratified survival patients into significantly different risk groups.Increasing α by over two orders of magnitude did not eliminate statistically significant survival stratification, whereas α = 1 significantly deteriorated classification performance.
- Comparative results: Federated models were generally competitive with centralized training and outperformed models trained within single institutions, but excessive noise created a privacy–performance trade-off.For survival prediction, only one of three single-institution models stratified patients into distinct survival groups, while site 3 performed comparably because it had a larger local dataset.
- Classification results: Federated learning achieved BRCA test AUCs of 0.927–0.932 and RCC micro-averaged test AUCs of 0.986–0.988 across tested random-noise levels.These results were evaluated on n=321 BRCA cases and n=170 RCC cases.
4. Conclusion
The framework demonstrates federated, attention-based weakly supervised learning for classification and survival prediction on gigapixel whole slide images without direct sharing of sensitive patient data. It may support more robust cross-institutional models while preserving differential privacy and enabling collaboration where slide transfer is constrained.
- The framework supports classification and survival prediction across whole slide images from different sites without direct sharing of potentially sensitive patient data.
- Federated learning may allow institutions to integrate whole slide image datasets and train models that generalize better to unseen data than single-institution models.
- Randomized mechanisms allow participating institutions to preserve differential privacy during federated model development.
- The approach could reduce barriers to cross-institutional collaboration, including when institutional or governmental regulations make transferring large quantities of slides impractical.
- Models that give institutions greater control over their data while achieving at or near state-of-the-art performance may support progress toward democratized computational pathology.