mDixon-based synthetic CT generation via transfer and patch learning

Xin Song, Pengjiang Qian, Jiamin Zheng, Yizhang Jiang, Kaijian Xia, Bryan Traughber, Dongrui Wu, Raymond F. Muzic

Research output: Contribution to journalArticlepeer-review

3 Scopus citations


We propose a practicable method for generating synthetic CT images from modified Dixon (mDixon) MR data for the challenging body section of the abdomen and extending into the pelvis. Attenuation correction is necessary to make quantitatively accurate PET but is problematic withPET/MR scanning as MR data lack the information of photon attenuation. Multiple methods were proposed to generate synthetic CT from MR images. However, due to the challenge to distinguish bone and air in MR signals, most existing methods require advanced MR sequences that entail long acquisition time and have limited availablity. To address this problem, we propose a voxel-oriented method for synthetic CT generation using both the transfer and patch learning (SCG-TPL). The overall framework of SCG-TPL includes three stages. Stage I extracts seven-dimensional texture features from mDixon MR images using the weighted convolutional sum; Stage II enlists the knowledge-leveraged transfer fuzzy c-means (KL-TFCM) clustering as well as the patch learning-oriented semi-supervised LapSVM classification to train multiple candidate four-tissue-type-identifiers (FTTIs); Stage III synthesizes CT for new patients’ mDixon images using the candidate FTTIs and voting principle. The significance of our method is threefold: (1) As the global model for patch learning, guiding by the referenced knowledge, KL-TFCM can credibly initialize MR data with overcoming the individual diversity. As the local complement, LapSVM can adaptively model each patch with low time and labor costs. (2) Jointly using the transfer KL-TFCM clustering and patch learning-oriented LapSVM classification, SCG-TPL is able to output accurate synthetic CT in the abdomen. (3) SCG-TPL synthesizes CT only using easily-obtainable mDixon MR images, which greatly facilitates its clinical practicability. Experimental studies on ten subjects’ mDixon MR data verified the superiority of our proposed method.

Original languageEnglish (US)
Pages (from-to)51-59
Number of pages9
JournalPattern Recognition Letters
StatePublished - Oct 2020

All Science Journal Classification (ASJC) codes

  • Software
  • Signal Processing
  • Computer Vision and Pattern Recognition
  • Artificial Intelligence


Dive into the research topics of 'mDixon-based synthetic CT generation via transfer and patch learning'. Together they form a unique fingerprint.

Cite this