System Model and Access Control Schemes for Medical Image Collaborative Analysis

Deep learning based medical image analysis has played an important role in the computer-aided diagnosis and treatment for diseases. The accuracy of classification has always been the primary goal pursued by researchers. However, the transmission process of images also faces the problems of limited bandwidth in WAN and increased risks of data security. Additionally, individual privacy is vulnerable when user data are exposed to an unauthorized user. To address these problems, this paper constructs a system model for collaborative analysis of diagnosis of diabetic retinopathy (DR). This model consists of two stages: data cleaning and lesion classification. In the data cleaning phase, the private cloud writes the trained model into the blockchain, other private clouds can use the best-performing model shared by private clouds on the blockchain to identify the image quality and transfer high-quality image to the lesion classification model for use. In the classification stage of lesions, each private cloud uses high-quality images for classification and uploads its model parameters to the public cloud for aggregation to obtain a global model. Then, the public cloud sends the global model to each private cloud to achieve collaborative learning, reduce the amount of data transferred, and protect personal privacy. The access control scheme includes the improved role-based access control (RAC) used within the private cloud and the blockchain-based access control scheme (BAC) used during the interaction between the private cloud and the public cloud. RAC can grant both functional and data access permissions to roles, and consider object attributes to realize fine-grained control. BAC is based on certificateless public key cryptography technology and blockchain technology, which can realize identity authentication and permission identification of private cloud while requesting to transfer model parameters from private cloud to public cloud, protect the identity, permission and model parameters of private cloud, and achieve lightweight access control. Two retinal datasets are utilized for the classification of DR. Experimental results demonstrate that data cleaning can efficiently remove low quality images and improve the accuracy of the classifica-tion for early lesions of DR. The accuracy is up to 90.2%.