Privacy-preserving medical image processing based on cryptography primitives: A review of methods

The application of machine learning (ML) in medical image processing has profoundly enhanced diagnostic efficiency for both healthcare providers and patients, particularly in intelligent oncology applications such as cancer screening, tumor segmentation, pathological image analysis, and precision treatment planning. However, the highly sensitive nature of medical and oncology imaging data mandates stringent security measures, making confidentiality a critical bottleneck for collaborative artificial intelligence in clinical practice. To address this challenge, the field is moving toward a deep convergence of security and ML. This paradigm aims to fully harness the analytical power of ML while strictly preventing the exposure or leakage of sensitive medical information. In this paper, we provide a comprehensive survey of privacy-preserving medical image processing frameworks based on cryptographic primitives. We not only review the architectural design of state-of-the-art schemes, including homomorphic encryption and secure multi-party computation, but also critically analyze their inherent limitations, particularly the trade-offs among computational overhead, model accuracy, and security guarantees, and discuss future research directions. The reviewed techniques provide a technical foundation for secure artificial intelligence in intelligent oncology, enabling privacy-preserving collaborative learning and encrypted inference across multiple medical institutions. Ultimately, this comprehensive review aims to facilitate the development of more robust, efficient, privacy-compliant, and trustworthy intelligent oncology systems.