Introduction
Low-Dose Computed Tomography (LDCT) is a critical medical imaging technique designed to minimize radiation exposure for patients. However, the reduction in dose inherently leads to increased noise and artifacts, which can compromise diagnostic accuracy.
While deep learning (DL) models have demonstrated remarkable success in denoising LDCT scans, their deployment in clinical settings faces a significant "privacy-utility" deadlock. Hospitals often lack the computational resources to train these models locally and are hesitant to upload sensitive patient data to third-party cloud services due to strict regulations like HIPAA and GDPR. Conversely, AI providers are reluctant to share their proprietary models to protect their intellectual property.
The Proposed Framework
This paper introduces a novel "plug-and-play" framework that enables standard deep learning models to perform denoising directly on encrypted data. By leveraging homomorphic encryption (HE), the system ensures that patient data remains confidential while the AI provider's model weights are never exposed to the hospital.
Mathematical Reformulation
The core of the approach lies in the use of Paillier Homomorphic Encryption, which supports additive and scalar multiplicative operations on ciphertexts. The researchers reformulated standard DL operations, such as convolutions and linear transformations, into a series of encrypted-domain operations. This allows the server to process encrypted images without ever decrypting them.
Handling Non-linearities
A significant challenge in encrypted-domain AI is the computation of non-linear activation functions like ReLU, which are not natively supported by HE schemes. To overcome this, the authors developed an interactive protocol:
- The server adds a random "perturbance matrix" to the encrypted features.
- The user decrypts these perturbed features and returns only the binary sign matrix (indicating positive or negative values).
- The server uses this sign information to apply the ReLU function losslessly.
Experimental Evaluation
The framework was evaluated using the 2016 NIH-AAPM-Mayo Clinic Low-Dose CT Grand Challenge dataset. The researchers simulated real-world conditions by adding synthetic Poisson and electronic noise to high-quality projection data.
Key Results
- Lossless Performance: The proposed method is mathematically proven to be lossless. The denoising results are identical to those achieved in the plaintext domain. For non-linear models, the framework achieved a Peak Signal-to-Noise Ratio (PSNR) of 40.37 dB and a Structural Similarity Index (SSIM) of 0.9789.
- Security: Attack simulations demonstrated that the random perturbation step effectively prevents users from reconstructing intermediate features, thereby protecting the server's intellectual property.
- Deployment: Unlike previous methods, this framework does not require model retraining or the use of non-standard activation functions, allowing for the seamless "wrapping" of existing high-performance models.
Conclusion and Future Work
This research provides a robust solution to the privacy challenges in medical AI. By enabling lossless denoising in the encrypted domain, it paves the way for secure cloud-based medical imaging services. Future work will focus on optimizing the framework for standard deep learning libraries like PyTorch and TensorFlow to improve computational speed and reducing communication overhead for deeper architectures.
Summary updated