Private Medical Image Segmentation on a Single Untrusted Server: A Homomorphic Encryption-Compatible Architecture and Its Measured Cost — code and data

The study runs medical image segmentation under pure, non-interactive CKKS homomorphic encryption with the Orion compiler, without bootstrapping, on two tasks: the lung field on chest radiographs (primary) and the lesion on breast ultrasound (BUS-BRA). This deposit holds the code that trained the networks and ran every encrypted measurement, the raw result files, the parameter sets, the weights of the two operating-point networks and the scripts that draw the data figures.Every table and figure value that comes from a measurement is contained in, or computed from, the files of this deposit. The image collections are needed only to retrain the networks or to repeat an encrypted run; they are public and are not redistributed here.