Resumen
Deep learning (DL)-based algorithms have demonstrated remarkable results in potentially improving the performance and the efficiency of healthcare applications. Since the data typically needs to leave the healthcare facility for performing model training and inference, e.g., in a cloud based solution, privacy concerns have been raised. As a result, the demand for privacy-preserving techniques that enable DL model training and inference on secured data has significantly grown. We propose an image obfuscation algorithm that combines a variational autoencoder (VAE) with random non-bijective pixel intensity mapping to protect the content of medical images, which are subsequently employed in the development of DL-based solutions. A binary classifier is trained on secured coronary angiographic frames to evaluate the utility of obfuscated images in the context of model training. Two possible attack configurations are considered to assess the security level against artificial intelligence (AI)-based reconstruction attempts. Similarity metrics are employed to quantify the security against human perception (structural similarity index measure and peak signal-to-noise-ratio). Furthermore, expert readers performed a visual assessment to determine to what extent the reconstructed images are protected against human perception. The proposed algorithm successfully enables DL model training on obfuscated images with no significant computational overhead while ensuring protection against human eye perception and AI-based reconstruction attacks. Regardless of the threat actor?s prior knowledge of the target content, the coronary vessels cannot be entirely recovered through an AI-based attack. Although a drop in accuracy can be observed when the classifier is trained on obfuscated images, the performance is deemed satisfactory in the context of a privacy?accuracy trade-off.