federated learning of healthcare for researching

this presentation contains the different papers

1. A Comprehensive Analysis of Recent Deep and Federated-Learning-Based Methodologies for Brain Tumor Diagnosis

2. Introduction This study examines the existing essential diagnostic methods for brain tumors and focuses on fundamental deep learning and federated learning methods using MRI images for brain tumor diagnosis. In addition, it provides a systematic analysis of the federated and deep learning literature on brain tumor detection and segmentation, while mainly focusing on classification. In recent years, much work has been performed on the automated diagnosis of brain tumors using deep learning, whereas only a few studies have been conducted on federated learning.

3. Cont.. Challenges with Deep Learning (DL): Deep learning methods often yield less accurate results when relying on smaller datasets for training and testing Solution: Federated Learning (FL): Federated learning is utilized to overcome the limited dataset issue by training a shared global model using data from several institutions without compromising patient data privacy

4. Research Method This study was conducted as a Systematic Literature Review (SLR) to categorize the state-of-the-art methods for brain tumor diagnosis

5. Cont.. Relevant articles were retrieved from 1 January 2017, to 20 December 2021 The databases included were Elsevier, ACM, IEEE, Springer, MDPI, Wiley, Miccai , and Medline Research not based on binary disease classification, research not using medical images for diagnosis, studies not identifying data sources, and articles based on non-human samples

6. Research Questions (RQs): What are the best available methods for the detection of a brain tumor? What are the metrics used to determine the performance of different methods used for brain tumor diagnosis? What datasets are used in recent research for the diagnosis of brain tumors? What is the quality of the selected papers? What is the impact of the selected papers on brain tumor detection?

7. Study Selection Outcome: Out of an initial 3986 identified studies, a total of 55 primary studies were finally included for the analysis

8. Data Analysis and Results RQ1: Methods for Brain Tumor Diagnosis: State-of-the-art techniques were categorized: Pretrained Classifiers: Architectures like CapsNets and U-Net with CNN. A major drawback is that their efficiency is often measured only on a single dataset Handcrafted Classifiers: These involve multiple steps (preprocessing, feature extraction, classification, segmentation). Examples include CNN combined with SVM (Support Vector Machine) or Watershed techniques

9. Cont.. Ensemble Classifiers: Combine multiple models (e.g., Deep Neural Network (DNN) and Probabilistic Neural Network (PNN)). The performance of ensemble classifiers generally works best on large datasets. Federated Learning (FL): Proposed to improve training by allowing collaboration between multiple organizations while protecting patient data privacy. FL uses DNN models for semantic segmentation of brain tumors

10. Cont.. RQ2: Performance Metrics: Classifier efficiency is determined by Sensitivity/Recall, Specificity, Precision, Accuracy, and Area Under the Curve (AUC). High accuracy results were reported, such as 99.63% accuracy achieved by a handcrafted + deep learning method on the BraTS 2016 dataset

11. Cont.. RQ3: Available Datasets: Data is categorized into publicly accessible benchmark datasets and non-public datasets Public Datasets: BraTS Challenge datasets (2012 through 2019, containing multimodal MRIs, often focusing on gliomas), Figshare Dataset (CE-MRI, containing 3064 T1-weighted images), TCGA-GBM (open-access), and ISLES (2015, 2017) Non-Public Datasets: These are generally available upon special request, such as the Combined Dataset (15,320 MRI images) and the BRAINIX Dataset

12. Cont.. RQ5: Impact: DL/FL systems can efficiently detect brain tumors and classify their types, helping radiologists improve their diagnostic accuracy. This includes multiclassification into categories like glioma, meningioma, and pituitary tumors

13. Discussion (Taxonomy and Model) Taxonomy for Brain Tumor Diagnosis: Tumors are classified into two main types: Primary Tumors: Develop from abnormal cell growth in the brain. They are categorized as benign or malignant. Common forms include Gliomas and Meningiomas. These can cause symptoms like vomiting, dizziness, and memory loss. Secondary Tumors (Malignant): These constitute the majority of brain cancers. They develop in other parts of the body (e.g., lung, breast, kidney, skin) and then spread to the brain.

14. Cont..

15. Common model for brain tumor

16. Cont.. Common Model for Brain Tumor Diagnosis: A five-step common model is proposed, based on the analysis of best-performing existing methods, to guide researchers: 1. Dataset Selection: Utilizing publicly accessible benchmark datasets and/or confidential data. 2. Data Analysis/Feature Selection: Applying fine-tuning to eliminate irrelevant details. 3. Identification of Training and Testing Images: Defining image sets properly (e.g., 75:25 ratio). 4. CNN-based Model Selection: Choosing a suitable technique (e.g., ensemble methods, handcrafted methods). 5. Finalization of the Best Model.

17. Open Issues and Challenges Data Imbalance: Datasets contain varying numbers of low-grade and high-grade tumors. • Tumor Size Detection: Identifying tumors where the lesion size is less than 3 mm remains challenging. • Patient Demographics (Age): Most data collected are from patients aged 40–70 years, indicating a lack of data from younger patients. • Need for Precision : The primary ongoing goal is to significantly increase sensitivity and improve the specificity and overall precision of the methods.

18. Principal Findings • Best Classifiers: Ensemble and handcrafted classifiers demonstrate outstanding results for brain tumor detection. • Optimal Data Split: The 75:25 ratio for training and testing sets generally provides the best accuracy for classifiers. • Most Used Datasets: The BraTS challenge and Figshare (CE-MRI) datasets are the most widely utilized by researchers.

19. Conclusions Deep learning and federated learning techniques produce better results compared to conventional methods. • DL Benefit: DL techniques reduce the need for complex and composite pre-processing techniques (like image resizing, cropping, and pixel normalization). • FL Benefit: FL successfully resolves the critical issue of using limited datasets for training and testing without compromising data privacy. • The wide diversity of available datasets makes it difficult to effectively compare and validate results across different studies. • Future research should focus on addressing data diversity, limited younger patient data, and enhancing model precision, sensitivity, and specificity.

20. A review on federated learning in computational pathology(2024)

21. Core Problem: Developing generalizable Computational Pathology (CPATH) algorithms requires large-scale, multi-institutional data. Healthcare data is subject to strict privacy rules, hindering the creation of large centralized datasets

22. Goal of the Review: To introduce core principles of FL. To provide an in-depth review of key developments of FL in CPATH

23. Basic Principles of FL Operational Workflow of FL (Horizontal Setup) 1. Initialization: Central server initializes and distributes a global model to clients. 2. Local Training: Each client trains the received model on its local, private dataset. 3. Model Update Transmission: Clients compute and transmit local model updates (gradients or weights) to the server. 4. Aggregation: Central server collects updates and aggregates them to form an updated global model. 5. Iteration: Steps 2-4 repeat (communication rounds) until the global model converges. 6. Deployment: The converged global model is deployed for inference, expected to generalize well to new data

24. Challenges and Considerations Data Heterogeneity (Non-IID data): Clients may have non-independent and identically distributed data, affecting global model convergence. • Privacy Leaks: Shared model parameters contain information about the data, requiring privacy preservation techniques. • Communication Overhead: Frequent communication between clients and the central server can lead to high costs. • System Heterogeneity: Varying computational capabilities among clients can cause delays (asynchronous updates). • Scalability: Less critical in CPATH (typically 2-100 clients) compared to consumer FL.

25. Study Design Research Question: • What are the key developments and findings of FL applications in the field of CPATH? Methodology and Scope: • Systematic review based on keyword searches (e.g., “federated” + “pathology”) on Google Scholar and Nature journals. • Timeframe: August 2021 – July 2024. • Inclusion Criteria: Studies applying FL methods for computational pathology, restricted to mammal-derived pathology data. Final Corpus: • A total of 15 studies were included for in-depth review.

26. CPATH Unique Challenges and PoCs Unique Challenges of CPATH 1. Gigapixel Image Size (WSIs): Whole Slide Images (WSIs) are massive (e.g., 150,000px x 50,000px) and require tile-based processing for DL models. Techniques like Multiple Instance Learning (MIL) are used for slide-level prediction using tile-level feature extraction. 2. Data Heterogeneity (Stain/Scanner Variation): Differences in tissue preparation, staining procedures, and scanners across labs lead to significant color and intensity variations. This challenge is directly amplified in a federated setting. 4.2. Proof of Concepts ( PoCs ) • Tasks Covered: PoCs demonstrate FL applicability across tasks like tile binary/multi-class classification (breast, colorectal, lung, kidney cancer), region segmentation, stain normalization, and slide-level prediction (e.g., Gleason grading, MSI prediction, survival prediction). • Evaluation: Comparisons are made against centralized models, local models, and ensemble models.

27. Key Performance Findings: FL models perform generally on par (±2%) or slightly worse (2–6% lower) than centralized models. ◦ FL models often show improved domain generalizability (> +2% or similar) compared to local models on external test sets. ◦ Most studies (excluding and) were conducted in simulated environments.

28. Technological Advances Technological advances primarily aim to address data heterogeneity and privacy preservation in CPATH FL.

29. Four Categories of Methods: 1.Model Aggregation Methods: Combining multiple client updates into a global model update. 2. Model Alignment Methods: Introducing additional loss objectives during local training to align local and global models. 3. Domain Alignment Methods: Addressing specific CPATH data heterogeneity (like staining and scanning variations). 4. Privacy Preservation Methods: Enhancing privacy to ensure sensitive data remains confidential

30. Tech Advances – Aggregation & Alignment Model Aggregation Methods • Standard Methods: Federated Averaging ( FedAvg ), FedProx , SCAFFOLD. • Personalized FL (PFL): Clients keep a subset of layers private (e.g., FedBN , FedPer ).

31. Cont.. Novel CPATH Methods: ◦ Prop-FFL: Rewards similar training loss across clients while accounting for sample proportion, reducing standard deviation in accuracy by 5%–11% compared to FedSGD . ◦ FedDropoutAvg : Randomly drops model parameters or client submissions during aggregation, acting as regularization and enhancing privacy. Outperformed FedAvg by 1–3% in F1-Score in a pilot study. ◦ SiloBN : A PFL method that keeps Batch Normalization (BN) parameters private, outperforming FedAvg by 1-2% AUC for breast tumor detection.

32. Cont.. Model Alignment Methods Introduce loss objectives to align local and global model feature representations. • Examples: FL Barlow Twins (uses contrastive learning on tile-level representations); training local models to reduce distance to a "global prototype"; using attention-consistency to ensure consistent identification of regions of interest. • Impact: Generally show only modest improvements (0–3% in accuracy)

33. Cont.. Domain Alignment Methods • Focus on mitigating staining and scanning variations. • GAN-based Stain Normalization: Training Generative Adversarial Networks (GANs) federatedly to generate more uniform stain appearances for downstream tasks. • Self-Supervised Learning (SSL): Used as a pre-training step, utilizing pseudo images generated by a multi-scale gradient GAN (MSG-GAN). • Impact: Demonstrated a substantially higher impact on performance (4% to >10%) compared to no normalization, highlighting the effectiveness of data-centric approaches.

34. Cont.. Privacy Preservation Methods • Differential Privacy (DP): Adds noise (e.g., Gaussian noise) to data or model updates to prevent exposure of individual data points. Low noise factors (z ≤ 0.1) resulted in slight performance degradation (up to -4% c-Index). • Secure Multi-party Computation (SMC): Masks individual updates using cryptographic techniques, allowing the server to see only aggregated updates. SMC showed less accuracy degradation (-0.5%) compared to DP (-6%) in one comparison.

35. Cont.. Opportunities for FL in CPATH: 1. Rare Cancers/Subtypes: FL enables model development where individual institutional datasets are sparse. 2. Multi-institutional Studies: Facilitates large-scale epidemiological studies and clinical trials while protecting sensitive patient data. 3. Weakly/Self-Supervised Learning (WSL/SSL): Combining FL with WSL/SSL empowers smaller institutions to pool weakly labeled or unlabelled data and computational resources.

36. Cont.. Implementation Challenges: • FL training time can be twice as long as centralized training. • Lack of standardized guidelines and software frameworks (e.g., NVFlare , Flower, TensorFlow Federated) complicates deployment. • Challenges in establishing a real-world FL ecosystem (e.g., system heterogeneity, setting up global server access within restrictive hospital networks) remain underexplored.

37. Revolutionizing healthcare data analytics with federated learning: A comprehensive survey of applications, systems, and future directions(2025)

38. Introduction The Data Challenge: In domains like healthcare, data is scattered across isolated silos due to stringent regulations (e.g., HIPAA, GDPR) and privacy concerns. Traditional centralized machine learning (ML) is often infeasible

39. How FL Works: 1. A central server shares an initial global model with clients. 2. Independent clients (e.g., hospitals) develop local ML models using their private data. 3. Clients relay the local model parameters (updates) back to the server, not the raw data. 4. The server aggregates the received models to create a new global model, which is then shared for further refinement. • Healthcare Impact: FL ensures the privacy of patients and healthcare providers, enabling collaborative analysis of raw data without explicit sharing.

40. Research Methodology Data Gathering: Literature research was conducted in multiple databases (Scopus, IEEE Xplore, JMIR, PubMed, and Web of Science) Search Scope: Articles published from January 2017 to January 2024 were targeted, primarily using keyword combinations of “federated learning” and “health *” for 2020–2024 Inclusion Criteria: The primary focus was on FL applications within healthcare settings that specifically foster data privacy and collaboration Exclusions: Articles introducing new FL techniques not explicitly designed for health-related applications were excluded

41. Selection Outcome: A meticulous selection process resulted in 159 studies being included in the final review. The review compares itself favorably to existing FL surveys by providing a deeper analysis of taxonomy, underlying ML models, privacy concerns, research gaps, and solutions

42. Critical analysis of existing FL surveys in healthcare

43. Related Literature (Distributed Learning Comparison)

44. Comparison of centralized and decentralized architecture and some of the recent applications.

45. Basic Building Blocks I (Architecture, Scale, Partitioning) This section covers the foundational elements for designing FL systems in healthcare

46. System Architecture Centralized: A central server orchestrates the entire training process. ◦ Pro: Simple, often adopted by small reliable devices (e.g., mobile devices). ◦ Con: Single point of failure. Most healthcare applications use this architecture. • Decentralized: Peer-to-peer communication replaces the central server, allowing clients to directly update global parameters. ◦ Pro : Reduces the risk of a single-point failure. Generally used by organizations (like hospitals). ◦ Con: Challenging to select a reliable leader for initiation and aggregation.

47. Data Availability Cross-Silo (Organizations/Hospitals): ◦ Scale: Relatively small number of clients (typically 2–100). ◦ Characteristics: High stability, strong hardware capacity, and reliable communication. ◦ Data: Usually non-independent and identically distributed (non-IID). • Cross-Devices (Mobile/IoT/Wearable Devices): ◦ Scale: Very large (millions of clients). ◦ Characteristics: Low stability; communication can be unreliable. ◦ Data: Usually IID .

48. Data Partitioning Horizontal Partitioning: Clients share identical attributes (features) but different participants (samples). Most FL-based medical applications adopt this approach. • Vertical Partitioning: Clients share the same participants (samples) but vary in attributes (features). Less widespread in medicine. • Federated Transfer Learning (FTL) Partitioning: Used when the first two methods are infeasible; involves some overlap in both sample and attribute spaces.

49. Machine Learning (ML) Models The choice of ML model depends on data availability, distribution, and client device capacities

50. Supervised Models Linear Models (LMs): Concise, straightforward, and robust (e.g., Logistic Regression, SVM). Extensively utilized for Electronic Health Record (EHR) prediction in FL, often implemented via Stochastic Gradient Descent (SGD). • Tree Models (TMs): Facilitate easy training and feature smaller model dimensions (e.g., Decision Trees, GBDTs). • Deep Learning (DL): Powerful Neural Networks (NNs, CNNs, RNNs) widely used for medical imaging, diagnostics, and time-series analysis. ◦ Challenge: Training deep NNs in FL requires substantially larger datasets and high computational resources, often resulting in prolonged training times and convergence issues.

51. Unsupervised Models Used when labeled data is scarce or unavailable. • Examples include clustering methods, K-Nearest Neighbor (K-NN), and Principal Component Analysis (PCA). FL clustering methods overcome the constraints of traditional centralized methods by handling large datasets while preserving privacy.

52. Privacy Techniques

53. Aggregation Purpose: To integrate local model parameters from all clients to generate a robust global model at each iteration. • FedAvg (Federated Averaging): The foundational and most widely adopted algorithm. Computes the global model as a weighted average of client weights based on their data sample size. ◦ Limitation: Convergence is hindered under non-IID data due to client drift and weight divergence. • FedProx : Modifies FedAvg by adding a proximal term to penalize local model deviation from the global model, improving robustness to data heterogeneity. • SCAFFOLD: Uses control variates to mitigate client drift, particularly effective under extreme non-IID settings and in resource-constrained environments.

54. Open-Source FL Systems FATE (Federated AI Technology Enabler): Considered the most comprehensive system. Supports numerous ML models (NN, DT, LM), horizontal/vertical partitioning, and multiple privacy mechanisms (DP, SMC, HE). • DataSHIELD : A framework tailored specifically for the domains of healthcare and social sciences. It enables advanced statistical analyses of individual-level data from multiple sources without data pooling.

55. Challenges and Open Problems in FL Designing efficient FL algorithms requires high accuracy, minimum processing durations, and reduced memory usage.

56. Communication Overhead Challenge: Distributed training requires intense communication across numerous clients over many iterations. • Solutions: Minimizing the total number of iterations, reducing communication load per iteration (e.g., sending updates from only a selected client subset), and reducing message sizes (e.g., feature selection, compression, or quantization). Compression techniques, however, may compromise model accuracy.

57. System Heterogeneity Challenge: Variations in client device capabilities (storage, power, network connectivity) can cause delays or dropouts. • Solutions: Implementing asynchronous communication protocols (e.g., FedAsync ), improving system fault tolerance, and employing lightweight models or model pruning

58. Non-IID Data (Statistical Heterogeneity) Challenge: Local datasets differ greatly in features and labels due to variations in geographic locations or EHR systems, leading to poor accuracy and slow convergence with standard FedAvg . • Solutions: Using clustered FL (grouping clients with similar distributions), employing meta-learning frameworks (e.g., MAML), or modifying aggregation algorithms (e.g., FedProx , SCAFFOLD). Data augmentation (GANs) can balance datasets but introduces potential privacy leakage concerns.

59. Attacks (Privacy and Security) Privacy Attacks: Adversaries exploit shared model updates to infer sensitive data characteristics. Examples include: ◦ Membership Inference: Determining if a specific sample was used for training. ◦ Property Inference: Inferring properties of the training subset (e.g., proportion of samples with a specific attribute). • Security Attacks (Poisoning): Malicious clients transmit erroneous or hostile updates, disrupting the global model. ◦ Data Poisoning: Malicious clients manipulate their local training data. ◦ Model Poisoning: Malicious parties alter local model updates or introduce backdoors. ◦ Byzantine Attacks: Nodes act maliciously, sending random or erroneous updates. Robust aggregation rules (Krum, trimmed mean) are employed to counter these.

60. FL for Healthcare Applications FL helps address data access challenges in precision medicine and enhances healthcare outcomes. • Data Modalities: FL systems handle diverse data types, including: ◦ Tabular Medical Data: EHR records (e.g., MIMIC-III, eICU Collaborative DB). ◦ One-Dimensional Data: Physiological signals (e.g., ECG signals). ◦ Image-Based Data: X-ray, CT, and MRI scans (e.g., COVID-19 diagnosis). • Primary Prediction Variables: The main FL-based predictions in the medical field include: ◦ Mortality Prediction (In-hospital, COVID-19 related). ◦ Adverse Drug Reaction (ADR) Prediction. ◦ Preterm Birth Prediction. ◦ Disease Diagnosis.

61. Cont … ML Models Used: There is no specific universal ML model for healthcare applications. Both linear models (LR, SVM) and deep learning models (NN, MLP, RNN) are widely utilized depending on the prediction variable. • Dominant System Components: Most studies focus on cross-silo architectures using horizontal data partitioning (as hospital EHR systems have identical attributes but diverse patient data). Differential Privacy (DP) and Homomorphic Encryption (HE) are the leading privacy techniques.

62. Summary and Outlook Ethical Necessity: The review emphasizes the necessity of ethical FL deployments, especially when dealing with sensitive medical data (e.g., mental health problems, HIV/AIDS, genetic conditions) where disclosure can cause severe social stigma. • Privacy Measures: FL-based healthcare systems must adopt strict privacy measures, such as DP (adding noise), HE (enabling encryption during computation), and SMC (patient confidentiality across institutions).

63. Major Obstacles and Future Directions Core Challenge: Designing an efficient and effective FL algorithm suitable for healthcare requires balancing competing factors: privacy strength, communication overhead, computation efficiency, and model accuracy

64. Privacy preservation for federated learning in health care