Abstract

Deep learning (DL) algorithms have significantly advanced healthcare applications, particularly in medical imaging for diagnosis, treatment planning, and disease management. However, the application of DL to sensitive medical images raises critical privacy and data security concerns. Balancing enhanced medical imaging analysis with the protection of patient confidentiality remains a formidable challenge. Consequently, privacy-preserving techniques for training and inferring DL models are increasingly vital. This study introduces a novel convolutional Bi-LSTM network designed to analyze fully homomorphically encrypted (HE) time-series medical images. Convolutional blocks extract selective spatial features from encrypted image sequences, while Bi-LSTM layers encode temporal dynamics. A weighted unit and sequence voting layer integrates spatial and temporal information with adaptive weighting to enhance diagnostic accuracy and minimize errors. The proposed framework is evaluated on two rigorous public benchmarks, CheXpert and BreaKHis, achieving an accuracy exceeding 0.99 on both datasets. These results surpass several competing methods, demonstrating the framework’s capability to effectively extract visual and sequential features from encrypted medical images while preserving patient privacy and delivering superior performance in medical image analysis.