Generative adversarial networks with adaptive normalization for synthesizing T2-weighted magnetic resonance images from diffusion-weighted images

• DOI: 10.3389/fnins.2022.1058487
• Publication Date: 2022-11-14T06:44:03Z
• Authors: Yanyan Mao; Chao Chen; Zhenjie Wang; Dapeng Cheng; Panlu You; Xingdan Huang; Baosheng Zhang; Feng Zhao
• Keywords: Normalization; Spatial normalization; Feature (linguistics); Feature vector

Recently, attention has been drawn toward brain imaging technology in the medical field, among which MRI plays a vital role clinical diagnosis and lesion analysis of diseases. Different sequences MR images provide more comprehensive information help doctors to make accurate diagnoses. However, their costs are particularly high. For many image-to-image synthesis methods supervised learning-based require labeled datasets, often difficult obtain. Therefore, we propose an unsupervised generative adversarial network with adaptive normalization (AN-GAN) for synthesizing T2-weighted from rapidly scanned diffusion-weighted (DWI) images. In contrast existing methods, deep semantic is extracted high-frequency original sequence images, then added feature map deconvolution layers as modality mask vector. This image fusion operation results better maps guides training GANs. Furthermore, preserve against common layers, introduce AN, conditional layer that modulates activations using fused map. Experimental show our method T2 perceptual quality detail than other state-of-the-art methods.