MRI to CT synthesis using contrastive learning
编号：51
稿件编号：62 访问权限：仅限参会人
更新：2021-10-30 17:37:08 浏览：1050次
张贴报告
摘要
Compared with CT, MRI enables more accurate delineation of target and organs-at-risk. But unlike CT, MRI images is not related with electron density for radiotherapy planning. The purpose of this paper is to generate pseudo CT for radiotherapy planning from MRI using deep learning method. Twenty-nine brain cancer patients with planning CT and diagnostic MRI were selected, of which 23 were used for training and 6 for testing. We use a new neural network based on contrast learning, called CUT. Meanwhile, we change the residual block to nine dense blocks and add a structural similarity to the loss function of the generator, the latter network is called denseCUT. We compare Hounsfield Unit (HU) discrepancies between pseudo-CT and original CT images. The mean absolute (MAE) errors were 72.0±6.9 HU，72.5±8.0 HU and 65.7±8.0 HU for the cycleGAN, CUT and denseCUT, respectively. Meanwhile, the structure similarity index (SSIM) were 0.91±0.01, 0.91±0.01 and 0.93±0.01, the peak signal-to-noise ratio (PSNR) were 28.5±0.7 dB, 28.5±0.7 dB and 29.4±0.8 dB, respectively. Experimental results show that the proposed denseCUT network is more accurate, robust, and efficient for predicting synthetic CT from MR images for MRI-only radiotherapy.
关键字
MRI, contrastive learning, pseudo CT
报告人
发表评论