Digital image processing. I. Evaluation of gray level correction methods in vitro.

Abstract

The aims of this study were a) to assess in an in vitro model the amount of density changes measured in digitally subtracted images due to electronic noise and image alignment error, and b) to test the accuracy of different gray level correction procedures in the reduction of densitometric image mismatches. A section of a pig mandible in which a hollow cylinder ITI Bonefit implant had been placed was used to obtain pairs of standardized radiographs. Series of radiographs were obtained with different exposure times (0.34, 0.39, 0.44, 0.51, 0.58 s). The radiographs were captured through a video camera, digitized and stored in a personal computer. The same radiographic image was recorded and subtracted from itself 10 times to study the error of the method due to electronic transformations of the images and image alignment. The noise due to the analog-to-digital transformation of the radiographic images was calculated to be +/- 2 gray levels i.e., 2% of the scale of gray levels. This kind of error was reduced up to 40% by capturing the images more than once and averaging the values per pixel. The manual superimposition of the images to be subtracted caused an increase of the error to +/- 3 gray levels (2.7%). Seven methods of gray level correction based either on a linear least squares approximation or on the cumulative density function (CDF) were tested. The group based on the CDF algorithm gave significantly better results than any other method. Pixels yielding differences smaller or equal to +/- 7 gray levels (5.5% of the scale of gray levels) should be excluded from further calculations in order to eliminate (false-positive) errors due to the normalizing algorithms. Furthermore, the CDF method on an arbitrarily chosen area of the image or on the wedge seems to give to subtraction images the ability of revealing real subtle changes in tissue density (fewer false-negative errors). The use of reference structures did not futher improve the ability of the normalization methods to correct gray level mismatches between radiographic pairs.