Encrypted login | home

Program Information

Bayesian MRI-Based Technique for Calculating Synthetic CT Images

no image available
M Simard

M Simard1*, A Pirenne1 , S Bedwani1,2 , D Blais3 , H Bouchard1,2 , (1) Universite de Montreal, Montreal, Quebec, (2) CRCHUM, Montreal, Canada (3) Centre hospitalier de l Universite de Montreal (CHUM), Montreal, Canada


MO-F-205-4 (Monday, July 31, 2017) 4:30 PM - 6:00 PM Room: 205

Purpose: To evaluate the potential of obtaining synthetic CT images from MRI sequences using a new Bayesian method

Methods: A novel method is founded from Bayes' theorem and the Gaussian mixture regression (GMR) model. This Bayesian GMR technique relies on the modelled likelihood of CT data given a measured MRI sequence and a prior distribution of CT numbers. Experimental data of three pig heads is acquired using a Siemens SOMATOM Definition Flash CT at 120 kVp and a Siemens MAGNETOM Aera MRI with four sequences: T1, T2, proton density (all three with Dixon contrast), and the ultra-short echo time PETRA. From the dataset, combinations of data are analyzed where each pig head is assigned either the role of sample (test), model (likelihood) or atlas (prior). Statistical analysis of the HU errors is performed separately for soft tissues and bones using an Otsu threshold in 1.5 million voxels per sample.

Results: Bayesian GMR yields an overall systematic improvement compared to the maximum likelihood GMR method. For the Bayesian GMR, the mean HU error ranges between -2 and 13 and the standard deviation ranges between 121 and 140, compared to the maximum likelihood GMR method yielding mean errors between -36 and 42 and standard deviations between 137 and 213. For soft tissues only, the Bayesian GMR method yields 68th percentile confidence intervals of HU errors of 8+/-27, 6+/-224 and 12+/-228 for the three samples. For bones only, the 68th percentile confidence intervals are of HU errors are found to be -73+/-127, 39+/-149 and 20+/-188.

Conclusion: Recent efforts in MRI-guided radiotherapy focus towards developing full MRI workflow. However, treatment planning requires accurate patient maps of electron density which are conventionally obtained from CT data. We propose a promising method to derive synthetic CT images from MRI sequences.

Contact Email: