Safe and Privacy-Aware AI Models for Medical Image Processing: A Multi-Paradigm Comparative Study

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

In addition to producing better models, hospitals that combine patient scans into a single training server also provide a single, alluring target: in 2023 alone, over 130 million healthcare records were compromised globally. This research begins with this contradiction between the confidentiality requirements of medicine and the data hungry of deep learning. We maintain the same architecture, preprocessing, hyperparameters, and evaluation protocol while training identical ResNet-50 classifiers on the NIH Chest X-ray corpus and a BraTS brain-tumor MRI set under three privacy regimes: Federated Learning, Differential Privacy, and Homomorphic Encryption. Federated Learning loses just 0.7 points (92.4%) while never sending raw images off-site, whereas centralized training achieves 93.1% accuracy. At ε = 1.0, Differential Privacy significantly reduces membership-inference attacks at a higher cost (89.7% accuracy). At the expense of about four times slower inference, homomorphic encryption maintains correctness at 92.8% and provides the highest confidentiality guarantee of the three. In bandwidth-tolerant multi-hospital contexts, Federated Learning provides the optimal accuracy-for-privacy exchange, while Differential Privacy and Homomorphic Encryption are better suited to institutions with lower latency budgets and stricter regulatory or secrecy limitations