An automatically computable fidelity measure which correlates well with observer preference is needed to facilitate the optimization of image processing methods. This study evaluates the use of the convolution mean squared error (CMSE) as such a measure. To compute the CMSE, both the true image and the "test image" are passed through a filter or other processing system. The mean squared error between the two identically processed images is then determined. A high-pass filter and, separately, a low-pass filter are optimized for this purpose. The inclusion of an early visual system model before these filters is also evaluated. The true image used was a high-resolution, high-count, nuclear medicine image of a liver and spleen phantom. Simulated acquisitions of this true image, which had been restored using the constrained least squares method with one of nine coarseness functions, provided the "test images." A low-pass filter of low cut-off frequency and low order gave CMSE values which correlated well (Spearman rank correlation coefficient (r_s) of 0.88) with average ranks from observer preference studies. A high-pass filter of high order and high cut-off frequency yielded similar results (r_s = 0.86).

Date Published: 1 January 1987
PDF: 9 pages
Proc. SPIE 0767, Medical Imaging, (1 January 1987); doi: 10.1117/12.967013

Published in SPIE Proceedings Vol. 0767:
Medical Imaging
Samuel J. Dwyer III; Roger H. Schneider, Editor(s)