Federated Learning-Based Privacy-Preserving Medical Image Diagnosis Using Explainable AI
- 1Chandigarh University
Abstract
Centralizing patient scans for training deep learning diagnostic models raises significant privacy and regulatory concerns, since medical images cannot easily leave the institution that produced them. This work presents a federated learning framework for pneumonia detection from chest radiographs in which multiple simulated hospital clients collaboratively train a shared convolutional neural network without transmitting any raw image data. A ResNet18 backbone, pretrained on ImageNet, is fine-tuned across five non-identically distributed clients using the Federated Averaging (FedAvg) algorithm implemented with the Flower framework. To address the opacity of the resulting model, Gradient-weighted Class Activation Mapping (Grad-CAM) is integrated so that every prediction is accompanied by a visual heatmap indicating the image regions that influenced the diagnosis. The system was trained on the Chest X-Ray Images (Pneumonia) dataset and evaluated against a centralized baseline trained on pooled data. The federated model reached 89.4% test accuracy compared with 91.5% for the centralized baseline, a gap of roughly two percentage points, while never exposing patient-level data to a central server. Interestingly, the federated model also showed more stable convergence across communication rounds than the centralized model exhibited across epochs, which we attribute to the implicit regularization effect of averaging weights from diverse local updates. A working prototype consisting of a FastAPI inference backend and a React-based diagnostic console was also built to demonstrate the end-to-end pipeline. The results indicate that federated learning combined with explainable AI can offer a practical, privacy-preserving alternative to centralized medical image analysis with only a modest reduction in diagnostic accuracy.
Related
- FedEAVE: federated evidence alignment and causal diagnosis for privacy-preserving personalized learning
- Causal-enhanced federated transfer learning framework for privacy-preserving cross-organizational strategic collaboration
- Federated learning with update level noise perturbation and blockchain audit logging for privacy-preserving healthcare ECG classification
- Federated learning for privacy-preserving opportunistic screening of lumbar-spine DXA-defined osteoporotic-level bone mineral density using chest radiographs: A multi-center independent validation study
- Privacy-preserving clustered federated learning via differential privacy and homomorphically encrypted prototypes
BibTeX
@misc{rahi2026federated,
title = {{Federated Learning-Based Privacy-Preserving Medical Image Diagnosis Using Explainable AI}},
author = {Shashi Ranjan Rahi and Shivangi Sharma and Harshwardhan Kumar},
howpublished = {Zenodo},
year = {2026},
month = oct,
doi = {10.5281/zenodo.23641248},
url = {https://zenodo.org/records/23641248},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Federated Learning-Based Privacy-Preserving Medical Image Diagnosis Using Explainable AI
AU - Shashi Ranjan Rahi
AU - Shivangi Sharma
AU - Harshwardhan Kumar
PY - 2026
DA - 2026/10/25
PB - Zenodo
DO - 10.5281/zenodo.23641248
UR - https://zenodo.org/records/23641248
AB - Centralizing patient scans for training deep learning diagnostic models raises significant privacy and regulatory concerns, since medical images cannot easily leave the institution that produced them. This work presents a federated learning framework for pneumonia detection from chest radiographs in which multiple simulated hospital clients collaboratively train a shared convolutional neural network without transmitting any raw image data. A ResNet18 backbone, pretrained on ImageNet, is fine-tuned across five non-identically distributed clients using the Federated Averaging (FedAvg) algorithm implemented with the Flower framework. To address the opacity of the resulting model, Gradient-weighted Class Activation Mapping (Grad-CAM) is integrated so that every prediction is accompanied by a visual heatmap indicating the image regions that influenced the diagnosis. The system was trained on the Chest X-Ray Images (Pneumonia) dataset and evaluated against a centralized baseline trained on pooled data. The federated model reached 89.4% test accuracy compared with 91.5% for the centralized baseline, a gap of roughly two percentage points, while never exposing patient-level data to a central server. Interestingly, the federated model also showed more stable convergence across communication rounds than the centralized model exhibited across epochs, which we attribute to the implicit regularization effect of averaging weights from diverse local updates. A working prototype consisting of a FastAPI inference backend and a React-based diagnostic console was also built to demonstrate the end-to-end pipeline. The results indicate that federated learning combined with explainable AI can offer a practical, privacy-preserving alternative to centralized medical image analysis with only a modest reduction in diagnostic accuracy.
ER -