Click here to


Are you sure ?

Yes, do it No, cancel

Fast Intensity Non-Uniformity Correction for MR Images Using Sparse Samples

L Shi*, S Perkins , C Moran , B Hargreaves , B Daniel , Stanford University, Stanford, CA


(Sunday, 7/29/2018) 4:00 PM - 4:55 PM

Room: Karl Dean Ballroom B1

Purpose: The MR images performed using small surface coils produce a higher signal-to-noise ratio but suffer from image non-uniformity artifacts, impeding its potential applications in adaptive radiotherapy treatment. In this work, we developed an effective and efficient intensity non-uniformity correction algorithm without pre-acquisition of the field sensitivity map or time-consuming post-filtration, two commonly used methods in commercial MRI systems.

Methods: Our method assumes two conditions: 1. The inhomogeneity field is smooth. 2. The acquired image is equal to the summation of the multiplication of the inhomogeneity field and inhomogeneity-free image and the noise. The method starts with a log transform on the image model where the multiplicative inhomogeneity field becomes additive, followed by a coarse tissue classification via a histogram-based method. The difference between classified tissues and their corresponding medians are considered as non-uniformity errors, but only sparse shading samples are used for correction to avoid classification errors. A Fourier-Transform based technique, local filtration, is used to efficiently process the sparse data for shading correction. The noise component is suppressed using edge-preserving adaptive diffusion filter. Method performance is evaluated on MR images for a breast phantom, 5 breast patients, 1 spine patient and 1 brain patient. Spatial non-uniformity (SNU) is used for evaluating non-uniformity suppression.

Results: The proposed method reduced the SNU from 244 arbitrary unit (AU) to 19 AU on breast phantom data, and an average SNU from 280 to 47 AU for breast patients, from 600 to 85 AU for spine patient and 280 to 21 AU for brain patient. Typical processing time for an image set of 512x512x199 takes about 3 seconds.

Conclusion: The proposed algorithm successfully removes image inhomogeneity for both phantom and patient MR images with negligible processing time. Such a method is readily implementable clinically as a software plug-in without needs of pre-acquisition calibration.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email