A multi-scale application of the N3 method for intensity correction of MR images

Spatial inhomogeneity due to the radio-frequency coil in MR imaging can confound segmentation results. In 1994, Sled introduced the N3 technique, using histogram deconvolution, for reducing inhomogeneity. We found some scans whose steep inhomogeneity gradient was not fully eliminated by N3. We created a multi-scale application of N3 that further reduces this gradient, and validated it on MNI BrainWeb and actual MRI data. The algorithm was applied to proton density simulated BrainWeb scans (with known inhomogeneity) and 100 standard MRI scans. Intra-slice and inter-slice inhomogeneity measures were created to compare the technique with standard N3. The slope of the estimated bias versus the known bias of BrainWeb data was 1.0 (r=0.9844) for N3 and 1.04 (r=0.9828) for multi-scale N3. The bias field estimated by multi-scale N3 was within 1% rootmean-square of that of standard N3. Over 100 MS patient scans, the average intra-slice measure (0 meaning bias-free) was 0.0694 (uncorrected), 0.0530 (N3) and 0.0402 (multi-scale). The average inter-slice measure (1 meaning bias-free) was 0.9121 (uncorrected), 0.9367 (N3) and 0.9508 (multi-scale). The multi-scale N3 algorithm showed a greater inhomogeneity reduction than N3 in the small percentage of scans bearing a strong gradient, and results similar to N3 in the remaining scans.