A Network for Medical Segmentation of Images with Multiple Centers That Preserves Privacy

An innovative system for safe medical image segmentation in multi-center collaborations, the Multi-Center Privacy-Preserving Network is introduced in this research. Reduce data transmission traffic while improving data privacy protection with our methodology’s novel method for multi-center collaborative learning. Our solution relies on a single encrypted data transfer, as opposed to federated learning’s round-by-round model data communication among the central server as well as local servers. The suggested CNN is a three-layer architecture that includes networks for encryption, segmentation, and decryption. Consequently, we provide a method for secure and adaptable cloud-based medical picture segmentation using a convolutional neural network. This research demonstrates that three cloud-based servers can train the algorithm using user-provided photos; nevertheless, they are unable to access comprehensive data on the model’s variables and user input. The experiment verifies that the technique, when put into effect, can guarantee the system’s security and efficiency while simultaneously reducing the user’s compute and storage loads. We use an encryption network to transform the raw picture data into ciphertext, and then we present a refined U-Net to separate the encrypted image data. The last step is to use a decryption network to get the segmentation mask. Computable image encryption is made possible by this design, which allows for ciphertext-based picture segmentation. We test our method on three datasets, one of which is a CTPA dataset, and two of which are cardiac MRI datasets. We show that MP-Net may safely use data from many sources to build a more accurate and comprehensive segmentation model.