Digital Image Processing Lab

Department of Radiology, University of Michigan

Research in Retrospective Correction of Image Intensity Inhomogeneities

MRI data sets are corrupted by multiplicative inhomogeneities, often referred to as nonuniformities or intensity variations, that hamper the use of quantitative analyses. We have developed an automatic technique that not only improves the worst corruptions such as those introduced by surface coils, but also corrects typical inhomogeneities encountered in routine volume data sets such as head scans without generating additional artifact. Because the technique uses only the patient data set, the technique can be applied retrospectively to all data sets, and corrects both patient independent effects such as rf coil design, and patient dependent effects such as tissue attenuation and dielectric-induced resonances experienced in high field MRI. Patient dependent attenuation effects are also encountered in x-ray computed tomography. All of the above are examples of multiplicative inhomogeneities which result in low spatial frequency corruption of acquired volume data sets. Our method uses an inhomogeneity-tolerant, preliminary segmentation technique followed by an estimate of the global corrupting polynomial (user-selected maximum variate order based on type of corruption) based on the preliminary segmentation followed by a simple correction of the original image using the estimated corrupting function. Following image correction, region of interest analyses, volume histograms, and thresholding techniques are more meaningful.

Rat brain surface coil MRI correction (304,291 bytes, b/w)
A rat brain surface coil MRI image acquired using a 7T magnet is shown in the upper left panel. The lower left panel illustrates the segmentation of the original image into 7 separate regions by the LCJ algorithm which provides the large region of support for computing the global B1 field inhomogeneity estimate, whose log is shown in lower right panel, from the original image. The corrected image shown in the upper right panel is computed by subtracting the log of the inhomogeneity estimate from the log of the original image and exponentiating. The line superimposed on the lower right inhomogeneity estimate shows the unity gain isocontour; thus in the correction, regions above the line are attenuated, and regions below the line are amplified by the correction. When the B1 field inhomogeneity estimate approaches 0 near the bottom of the data set, its reciprocal is very large which leads to the noisy white region at the bottom of the corrected image. This correction results from the use of a mvo=5 polynomial model. Model orders of 4 and 5 seem necessary for surface coil images, while model orders of 2 are appropriate for standard head acquisitions (bird cage rf coils: body coil transmit, and head coil receiver).

Head bird cage coil MRI correction (30,110 bytes, color)
Axial MRI data sets acquired using a 1.5 T GE Signa using a "bird cage" rf head coil; patient has a large left hemisphere infarct. Three rows represent superior (top row), midbrain (middle row, and base of cerebellum (lower row). Columns left of center are associated with original data acquisition, and columns right of center are associated with the corrected data. The central column displays the difference between the two adjacent columns: immediately to the left is the uncorrected, acquired, T1-weighted image set, while immediately to the right of center is the corrected T1-weighted set. The window/level settings for the three central gray scale columns are exactly the same, with the exception that the level of the difference images has been adjusted such that zero difference yields black (cursor, central image). A narrow window was chosen to emphasize differences. The outermost two columns display cluster labels corresponding to "gray matter" (gray) and CSF (white) obtained from a global trivariate, histogram-based clustering (HICAP) of the uncorrected (left-most column) and corrected (right-most column) data volumes.

Breast flex coil intensity correction (54,997 bytes, b/w)
Images obtained using a flexible surface coil for MRI breast imaging on a 1.5 T, GE Signa scanner. Uncorrected (left column), LC segmented (middle column), and corrected (right column) images are presented at a superior level of the breast (top row), midbreast (middle row), and inferior level of the breast (bottom row). Images in left and right columns are presented at identical contrast settings.

Breast flex coil intensity correction with region merging (40,133 bytes, color)
Images obtained using a flexible surface coil for MRI breast imaging on a 1.5 T, GE Signa scanner. The left column is from the original data set, the middle column is the result of the correction shown in the figure above, and the right column is an improved correction as the result of merging clusters in the modeling process. Gray scale images (top row) and corresponding grayscale coded uniformly thresholded version (bottom row). In the original LCJ algorithm the number of separate segments was determined by spatial separation, i.e. each volume hypothesis that passed the uniformity check was assigned a separate volume identification number for use in computing the correction polynomial. Realizing that some larger volumes are artificially divided into smaller ones by the segmentation process, we noted that the stability of the correction algorithm could be improved by merging segments corresponding to the same tissues and thus reducing the degrees of freedom for the same mvo model. A retrospective merging of segments was performed, followed again by computation of the correction polynomial. The improved uniformity obtained in both breasts is obvious in the right column's images.

Research by C.R. Meyer, P.H. Bland and J.G. Pipe.

(For related abstracts see the IEEE TMI abstract or the CVRMed95 abstract .)

This work was supported in part by DHHS PHS grants NIH 1R01CA52709 and 1F32CA09309.