Privacy-Preserving Lung Disease Classification via ResNet50 Feature Extraction and Fully Homomorphic Encryption

Maintaining patient privacy while ensuring diagnostic accuracy poses a critical challenge in the advancement of AI applications in healthcare. This study proposes an approach to lung disease classification leveraging Fully Homomorphic Encryption (FHE) through the Concrete-ML framework, enabling secure inference on encrypted data without exposing sensitive medical information. We evaluate the performance of FHE-based inference using two machine learning models: Logistic Regression (LR) and Decision Tree (DT). When applied to encrypted features extracted from chest X-rays, these models achieved classification accuracies of 73% and 51%, respectively. Given the computational overhead associated with FHE, direct processing of raw images and complex models remains impractical. To mitigate this, input images are pre-processed–typically resized using bicubic interpolation–and feature extraction is conducted using convolutional neural networks such as ResNet50 prior to encryption. Results show that while Decision Tree are limited under FHE, Logistic Regression offers a practical balance between privacy and performance.