One-Sided Unsupervised Medical Image Synthesis with Normalized Edge Consistency

Vincent Jaouen, Pierre-Henri Conze, Dimitris Visvikis

Abstract: We introduce a novel one-sided approach for unsupervised image-to-image (I2I) translation, referred to as Normalized Edge Consistency (NEC), to address some limitations of methods like CycleGAN and CUT, which can produce realistic images at the cost of limited faithfulness (or structural consistency) with respect to the source. Inspired by the Normalized Gradient Field similarity used in image registration, NEC incorporates a normalized gradient vector field loss combined to an adversarial objective to maintain attachment to local orientation of structures while allowing for realistic contrast changes. NEC is easy to implement and is governed by a unique parameter that controls its sensitivity to noise. We demonstrate NEC’s (suprising) efficacy in T1-toT2 brain MRI translation using unpaired data from the BraTS 2023 dataset, rivaling with supervised methods like Pix2Pix.