Source-linked AI summary
Collaborative Federated Learning For Healthcare: Multi-Modal COVID-19 Diagnosis at the Edge
Adnan Qayyum, Kashif Ahmad, Muhammad Ahtazaz Ahsan, Ala Al-Fuqaha, Junaid Qadir
TL;DR
Cloud-based healthcare can face bandwidth, latency, transmission-cost, security, and privacy challenges, especially for remote centers lacking advanced diagnostic facilities. The paper proposes clustered federated learning for privacy-aware multimodal COVID-19 diagnosis at the edge and reports performance comparable to specialized federated models while outperforming conventional multimodal federated learning.
Problem
Remote healthcare centers may lack advanced diagnostic facilities, while cloud-based infrastructures face bandwidth, latency, transmission-cost, security, and privacy challenges.
Method
The paper uses clustered federated learning to train a collaborative multimodal COVID-19 model on X-ray and ultrasound data without sharing local data.
Results
The CFL model achieves results comparable to specialized federated models and performs considerably better than a conventional multimodal federated model.
Takeaways & Limitations
CFL can help remote healthcare centers benefit from collaborative learning while processing multimodal clinical visual data at the edge without sharing local data.
Takeaways & Limitations
Deploying deep learning models at the edge is constrained by power-hungry computation, limited memory, battery availability, and heterogeneous system resources.
Abstract
from arXiv · showhide
Despite significant improvements over the last few years, cloud-based healthcare applications continue to suffer from poor adoption due to their limitations in meeting stringent security, privacy, and quality of service requirements (such as low latency). The edge computing trend, along with techniques for distributed machine learning such as federated learning, have gained popularity as a viable solution in such settings. In this paper, we leverage the capabilities of edge computing in medicine by analyzing and evaluating the potential of intelligent processing of clinical visual data at the edge allowing the remote healthcare centers, lacking advanced diagnostic facilities, to benefit from the multi-modal data securely. To this aim, we utilize the emerging concept of clustered federated learning (CFL) for an automatic diagnosis of COVID-19. Such an automated system can help reduce the burden on healthcare systems across the world that has been under a lot of stress since the COVID-19 pandemic emerged in late 2019. We evaluate the performance of the proposed framework under different experimental setups on two benchmark datasets. Promising results are obtained on both datasets resulting in comparable results against the central baseline where the specialized models (i.e., each on a specific type of COVID-19 imagery) are trained with central data, and improvements of 16\% and 11\% in overall F1-Scores have been achieved over the multi-modal model trained in the conventional Federated Learning setup on X-ray and Ultrasound datasets, respectively. We also discuss in detail the associated challenges, technologies, tools, and techniques available for deploying ML at the edge in such privacy and delay-sensitive applications.
I. INTRODUCTION
The paper motivates edge-based healthcare learning by combining privacy, security, and low-latency requirements with collaborative COVID-19 diagnosis for remote centers. It proposes clustered federated learning for multi-modal X-ray and Ultrasound imagery without sharing local data.
- Motivation: Edge computing brings storage and computation closer to users, reducing private-data exchange with centralized servers and supporting healthcare’s security, privacy, reliability, and latency requirements.These requirements are especially important where remote centers lack advanced diagnostic facilities and cloud infrastructure faces bandwidth, latency, transmission-cost, and privacy challenges.
- Motivation: Remote healthcare centers can use telemedicine to transmit ultrasound and X-ray images for diagnosis, but cloud-based infrastructures introduce bandwidth, latency, cost, and privacy challenges.Edge computing stores and processes data locally or closer to edge devices, resulting in lower latency and increased privacy and data security.
- Research Aim: The paper analyzes intelligent edge processing of COVID-19 clinical visual data so remote centers can participate in multi-modal collaborative learning without sharing local data or its modality information.The authors identify relatively little prior work focused on edge AI for COVID-19 diagnosis.
- Approach: The proposed CFL framework trains a single model to diagnose COVID-19 separately in X-ray and Ultrasound images using visual data collected across centers.The paper defines a multi-modal model as one model capable of diagnosing both imagery types when provided separately.
- Evaluation and Scope: The paper evaluates CFL on two benchmark datasets, compares it with specialized and conventional multi-modal federated-learning baselines, and discusses edge-deployment challenges and enabling technologies.Its contributions include studying distribution divergence between X-ray and Ultrasound data and related open research issues.
II. BACKGROUND
The background reviews automated COVID-19 diagnosis across CT, Ultrasound, and X-ray imaging, alongside federated-learning approaches intended to preserve healthcare-data privacy. It also identifies limited CT data and the resource demands of edge deployment as important challenges.
- Existing Automated COVID-19 Diagnosis Work: Automated COVID-19 image analysis mainly uses CT, Ultrasound, and X-ray modalities, supported by publicly available datasets.These efforts aim to speed diagnosis through image-based methods.
- Data and Model Constraints: Limited availability of large-scale CT datasets motivates alternative architectures and transfer learning, including Capsule Networks for X-ray diagnosis and pre-training on 94,323 chest X-ray images.The cited COVID-CAPS approach uses an external dataset for pre-training and transfer learning.
- X-ray-Based Methods: X-ray diagnosis has also been studied with fine-tuned deep models, SVM classifiers using ResNet-50 features, and evaluations of several existing deep architectures.These approaches classify COVID-19 and non-COVID-19 cases or assess model performance on X-ray imagery.
- Federated Learning: Federated-learning studies collaboratively train models across hospitals or sources to support privacy-preserving COVID-19 diagnosis, including approaches based on CT scans and blockchain-FL.The reviewed work establishes federated learning as part of image-based COVID-19 diagnosis research.
1) Resource Scarcity and Heterogeneity:
Deploying deep learning at the edge is constrained by limited device resources, heterogeneous connectivity and data distributions, privacy threats, and adversarial attacks.
- Resource Scarcity and Heterogeneity: Deep learning models are computationally expensive and difficult to deploy on low-power edge devices with limited memory.Training and inference typically rely on power-hungry GPUs and servers, whereas edge devices operate under frugal power and memory budgets.
- Network Communication: Heterogeneous computational and communication resources can produce slow, unstable communication and disconnect unstable clients.Reliable, powerful Internet connectivity is desirable when models are trained at client sites; edge deployment nevertheless avoids transmitting raw local data to the server.
- Statistical Heterogeneity: Statistical heterogeneity across device-generated datasets complicates optimization and hyperparameter training, affecting model performance.Device-specific modeling, including meta-learning, is identified as one possible response.
- Privacy and Security Challenges: Malicious devices can expose sensitive information from shared model parameters even without direct access to the underlying data.Cryptographic approaches and differential privacy are cited as privacy-preserving techniques.
- Adversarial ML: Compromised clients can launch poisoning attacks in federated learning, increasing error rates of distributively trained models.The paper also notes broader vulnerability to carefully crafted adversarial examples.
C. Enabling Technologies: Building Blocks for ML at Edge
Edge ML enables local inference and collaborative processing while using sensor fusion, transfer learning, generative models, and self-improving devices to expand edge capabilities.
- Enabling Technologies: Edge devices can use ML inference to predict unseen objects or classes, reducing latency and bandwidth in telemedicine.Examples include cancer prognosis, brain tumor classification, and other clinical data analysis.
- Enabling Technologies: Sensor fusion combines information from multiple sensors or sources to use available data more efficiently and solve complex problems more accurately.Healthcare applications can combine several sensor sources to predict clinical events.
- Enabling Technologies: Transfer learning reuses knowledge from one domain through fine-tuning a pretrained model, supporting smaller datasets with fewer computational resources.In healthcare IoT, pretrained models shared from the cloud can help balance workload and latency.
- Enabling Technologies: Generative models can approximate original cloud data for edge-device training when samples are limited or tasks are computationally complex.Generative deep models have also been explored for synthetic medical-image generation.
- Enabling Technologies: Self-improving devices can optimize performance continuously using collected data and the behaviors of other devices.These strategies support faster device configuration and more efficient implementation and deployment.
2) Hardware Optimization Techniques:
Hardware optimization for edge ML combines decentralized computing and specialized processors with resource-efficient model architectures, pruning, and distillation.
- Hardware Optimization Techniques: Successful edge ML deployment requires both low power consumption and high performance, motivating custom hardware architectures.The paper discusses hardware optimization techniques for reducing resource demands.
- Hardware Optimization Techniques: Blockchain and smart contracts can support decentralized sharing of computational resources among end and edge devices.This approach helps eliminate reliance on centralized entities in decentralized edge-computing ecosystems.
- Hardware Optimization Techniques: AI co-processors accelerate edge ML through parallel computing with multiple mini-cores and distributed-memory data fetching.Specialized processors can be integrated into heterogeneous systems optimized for particular tasks.
- Algorithmic Optimization Techniques: Parameter-efficient networks, such as MobileNet and SqueezeNet, reduce computation costs for training and inference on embedded devices.These architectures are designed for computation- and memory-efficient ML/DL deployment.
- Algorithmic Optimization Techniques: Network pruning removes neurons that do not contribute to final accuracy, saving memory while retaining model performance.Google’s Learn2Compress reportedly reduced neurons by a factor of 2 while retaining 97% overall accuracy.
- Algorithmic Optimization Techniques: Network distillation transfers knowledge from a larger model to a smaller model, reducing model size without substantially compromising accuracy.The method is described as useful for resource-constrained IoT platforms.
III. COVID-19 DIAGNOSIS USING COLLABORATIVE FEDERATED LEARNING
The paper formulates multi-modal COVID-19 diagnosis as collaborative learning with clustered federated learning, separating X-ray and ultrasound clients while aggregating their updates into a shared model.
- Problem formulation: CFL learns a shared model Ms from disjoint client datasets grouped into modality-specific clusters with different distributions Dk.Cluster C1 contains X-ray facilities and cluster C2 contains ultrasound facilities; clients do not know other clients’ identities or data.
- Collaborative training: Each client trains the shared model locally on labeled samples and sends learned weights Wr,k,m to the server after each communication round.The server distributes initial weights W0 and the shared model to clients before local training.
- Aggregation and evaluation: The server performs federated averaging on client weight updates, updates Ms, evaluates it on merged X-ray and ultrasound test data, and repeats the process.The merged test set provides multi-modal evaluation at each communication round.
- Algorithm: Algorithm 1 iterates local training and weight collection across communication rounds using specified clients, epochs, batch size, and learning rate.The algorithm initializes W0 and repeats the procedure for r = 1, ..., R.
B. Experimental Setup
The experiments use chest X-ray and ultrasound datasets for binary COVID-19 classification, with separate training and testing splits and a merged multi-modal test set. The datasets contain substantial variability and class imbalance, complicating generalized-model training.
- Data description: The study uses two datasets from different sources containing chest X-ray and chest ultrasound images for binary classification of COVID-19 versus normal cases.Each dataset is split into 80% training and 20% testing data.
- Data distribution: The server evaluates the shared model using a joint testing set formed by merging the X-ray and ultrasound testing sets.Training data are distributed among clients according to the number of clients in each modality cluster.
- Dataset characteristics: The datasets exhibit inter- and intra-class variability in image size, quality, contrast, brightness, and subject positioning.Figure 3 illustrates these variations across the X-ray and ultrasound datasets.
- Dataset limitations: The datasets are highly imbalanced, and their variability and nonstandard curation make training a generalized model more difficult.The databases were curated from different public sources rather than standardized for COVID-19 detection.
2) Model Architecture and Implementation Details:
The implementation uses an extended VGG16-based model with grayscale 256×256 inputs and Adam optimization. Focal loss addresses class imbalance, while experiments run in a simulated TensorFlow environment.
- Model architecture: The model extends VGG16 with one additional convolutional layer and three fully connected layers before the original output layer.Images are converted to grayscale, resized to 256×256, and normalized before inference and training.
- Optimization: The model is trained at each client with Adam using a learning rate of 0.0001.The implementation also applies preprocessing before images are fed into the model.
- Loss function: Focal loss addresses imbalanced binary classes by adding the modulating factor (1 − pt)^γ to standard cross-entropy loss.Its α-balanced variant uses α to balance the importance of positive and negative examples, with γ ≥ 0 as a tunable focusing parameter.
- Implementation environment: The proposed work is implemented with TensorFlow, and all experiments are performed in a simulated environment.
IV. RESULTS AND DISCUSSIONS
CFL provides multimodal COVID-19 diagnosis with performance comparable to specialized federated models and better than conventional multimodal FL. The experiments also examine communication-round behavior and identify privacy, efficiency, and optimization considerations.
- Performance comparison: Precision, recall, and F1-score were used to compare CFL with specialized FL and conventional multimodal FL on X-ray and Ultrasound imagery.The experiments report both per-class and overall results because the datasets are imbalanced.
- Performance comparison: CFL achieved overall results comparable to specialized models trained separately for X-ray and Ultrasound imagery, while outperforming conventional multimodal FL.The evaluation used specialized FL and multimodal conventional FL as baselines.
- Communication-round behavior: CFL maintained comparable performance with specialized FL across increasing communication rounds, whereas conventional multimodal FL overfitted after 50 rounds.CFL also showed overfitting at later communication rounds, while specialized-model parameters diverged beyond the marked inflection point.
- Lessons learned: CFL enables multimodal collaborative learning without sharing local data or explicit information about the data modality.The approach aggregates model information rather than centralizing clinical images.
- Lessons learned: Compared with conventional FL, CFL provides better performance when data distributions diverge across modalities.The reported divergence can involve both class distributions and the nature of the image samples.
- Limitations and challenges: The framework faces efficiency, security, and CFL-parameter optimization challenges, alongside a performance trade-off between centralized and distributed training.The discussion also highlights the need for early stopping and improved stopping criteria.
A. Developing Personalized Approaches
Edge healthcare learning must accommodate heterogeneous clients, delayed participation, and security threats. CFL addresses source-related distribution divergence, while broader data-distribution variation remains a future evaluation target.
- Edge clients differ in data, acquisition, network, computation, and storage resources, creating substantial heterogeneity.
- Adversarial defenses remain attack-specific and may fail against unforeseen attacks, leaving robust ML/DL models an open problem.
- Delayed data acquisition or inactive clients can delay federated parameter updates and affect overall system performance.
- CFL better copes with distribution divergence between X-ray and Ultrasound sources, but variation in samples per client remains unevaluated.