James B. Arnold², Jeih-San Liow⁴, Kirt A. Schaper², Joshua J. Stern³, John G. Sled⁵, David W. Shattuck⁶, Andre J. Worth⁷, Mark S. Cohen⁸, Richard M. Leahy⁶, John C. Mazziotta⁸, David A. Rottenberg^1,2,3,4

(1) Neurology Service and (2) PET Imaging Center, Minneapolis VA Medical Center
Departments of (3) Neurology and (4) Radiology, University of Minnesota
(5) McConnell Brain Imaging Centre, Montreal Neurological Institute
(6) Signal and Image Processing Institute, University of Southern California
(7) Center for Morphometric Analysis, Department of Neurology, Massachusetts General Hospital
(8) Brain Mapping Center, Neuropsychiatric Institute, and Department of Neurology, UCLA School of Medicine

ABSTRACT

The desire to correct intensity non-uniformity (bias) in magnetic resonance imaging (MRI ) data sets has led to the proliferation of non-uniformity-correction algorithms with different theoretical underpinnings. In order to provide end users with a rational basis for selecting a given algorithm for a specific neuroscience application, we evaluated the performance of six non-uniformity-correction algorithms using simulated and real MRI data volumes. The addition of known biases to a simulated MRI volume (the Montreal Brain Phantom) allowed for direct measurements of the accuracy of bias removal. Because simulated data do not adequately reflect the signal, noise and bias characteristics of real image data, algorithm performance was also evaluated on a set of six repeat T1-weighted MRI scans of a single subject acquired over a six-month period using the same 1.5T scanner; in this way, it was possible to quantify the reproducibility of bias removal. We also compared algorithm performance on data volumes of different subjects from different (1.5T and 3.0T) scanners in order to relate differences in performance to inter-subject variability and/or differences in scanner performance. Finally, we applied each algorithm iteratively to its own output to assess the stability of the initial bias estimate and to identify any tendency for the algorithm to introduce noise into the test volumes. In phantom studies, the correlation of the extracted with the applied non-uniformity was highest in the transaxial (left-to-right) direction and lowest in the axial (top-to-bottom) direction.  Only two of the tested algorithms demonstrated a high degree of stability, as measured by the iterative application of the algorithm to its corrected output. While none of the algorithms that we evaluated performed ideally under all circumstances, locally adaptive methods generally outperformed nonadaptive methods.

ABBREVIATED RESULTS

- Biased Brain Phantom Data (Figs. 1-4)
- Multiple Repeat T1-weighted MRI Scans of a Single Subject (Figs. 5-10)
- High-Resolution T1-weighted MRI Scans from Two Different Centers (Figs. 11-12)

DATASETS

Phantom Studies

unbiased, unnoised brain phantom (3.8MB)		brain mask (67KB)
+/- 2% paraboloidal bias field (23MB)		+/- 4% paraboloidal bias field (24MB)
+/- 8% paraboloidal bias field (25MB)		+/- 2% sinusoidal bias field (22MB)
+/- 4% sinusoidal bias field (22MB)		+/- 8% sinusoidal bias field (23MB)

Repeat MRI Scans

repeat #1 (2.7MB)		brain mask #1 (40KB)
repeat #2 (2.7MB)		brain mask #2 (40KB)
repeat #3 (2.7MB)		brain mask #3 (40KB)
repeat #4 (2.7MB)		brain mask #4 (40KB)
repeat #5 (2.7MB)		brain mask #5 (40KB)
repeat #6 (2.7MB)		brain mask #6 (40KB)

High Resolution MRI's

UMN MRI #1 (6.8MB)		UMN brain mask #1 (74KB)
UMN MRI #2 (7.2MB)		UMN brain mask #2 (81KB)
ICBM MRI #1 (4.1MB)		ICBM brain mask #1 (55KB)
ICBM MRI #2 (3.9MB)		ICBM brain mask #2 (57KB)

 

SOFTWARE

ftp://neurovia.umn.edu/pub/bias_correction/programs.tgz (32KB)

© 2000 by the International Neuroimaging Consortium with funding from the  Human Brain Project and by the Regents of the University of Minnesota.  Contact: webmaster@neurovia.umn.edu

The University of Minnesota is an equal opportunity educator and employer.