NVIDIA and the Royal College of Physicians Launch Patient Privacy Protection System

2019-10-21

In order to protect patient privacy data and promote medical research, NVIDIA researchers have collaborated with researchers from the Royal College of Medicine in the UK and the French company Owkin to launch the first privacy-preserving learning system for medical image analysis, providing joint learning for the UK Medical Imaging and AI Healthcare Center.


Recently, at the top-level conference MICCAI on medical imaging held in Shenzhen, they published a paper on this research. NVIDIA and researchers from the Royal College of Medicine in the UK detailed how they developed their technology.Researchers said in the paper: "Federated learning is a learning paradigm that allows developers to train a centralized deep neural network (DNN) using training data distributed across multiple locations. Federated learning enables collaborative and decentralized training of neural networks without sharing patient data. Each node trains its own local model and periodically submits it to the parameter server. The server accumulates and aggregates data from various parts to generate a global model, which is then shared with all nodes."


Researchers explained that although federated learning can provide a high level of security in terms of privacy, there are still methods to reconstruct data through model inversion. To help make federated learning more secure, researchers studied a method called ε-differential privacy framework (a formal method for defining privacy loss), which protects patient and institutional data with a powerful privacy protection system. To prioritize patient privacy, Owkin has built differential privacy and other state-of-the-art privacy protection technologies into the architecture.


This breakthrough experiment was conducted based on the brain tumor segmentation data from the BraTS 2018 dataset. The BraTS 2018 dataset contains MRI scans of 285 brain tumor patients. The dataset is used here to evaluate federated learning algorithms on multimodal and multiclass segmentation tasks. The team then adopted the state-of-the-art training pipeline originally designed for training in the dataset as part of the NVIDIA Clara Train SDK.To perform training and inference, the team used the NVIDIA V100 Tensor Core GPU. When compared with data-centric systems, the proposed method achieves comparable segmentation effects without the need to share institutional data.


Furthermore, the experimental results indicate a natural correlation between privacy protection and the quality of trained models. Nevertheless, using sparse vector technology, the federated learning system can still provide strict privacy protection at a relatively low cost in terms of model performance.


The NVIDIA team explained that deep learning is a powerful technique for automatically extracting knowledge from medical data. Federated learning has the potential to effectively aggregate knowledge learned locally from private data across different institutions, thereby further improving the accuracy, robustness, and generalization ability of deep models.


NVIDIA researchers stated, "This research is an important step towards deploying secure federated learning, which will make large-scale data-driven precision medicine possible."

RECENT AI NEWS

RECENT AI TOOLS