Frequency-Supervised MR-to-CT Image Synthesis

This paper strives to generate a synthetic computed tomography (CT) image from magnetic resonance (MR) image. The CT is valuable for radiotherapy planning when only an MR available. Recent approaches have made large strides in solving this challenging synthesis problem with convolutional neural networks that learn mapping inputs outputs. In paper, we find all existing share common limitation: reconstruction breaks down and around the high-frequency parts of images. To address limitation, introduce frequency-supervised deep explicitly enhance MR-to-CT reconstruction. We propose frequency decomposition layer learns decompose predicted outputs into low- components, refinement module improve through adversarial learning. Experimental results on new dataset 45 pairs 3D MR-CT brain images show effectiveness potential proposed approach. Code available at https://github.com/shizenglin/Frequency-Supervised-MR-to-CT-Image-Synthesis.