Channelized hotelling and human observer correlation for lesion detection in hepatic SPECT imaging

Mathematic "model" observers that predict human performance are of interest in medical imaging as substitutes in psychophysical studies. We have exami ned the correlations between human observers and several forms of the chann elized Hotelling observer (CHO) for a tumor detection task with simulated S PECT liver images that were used to study the effects of scatter and scatte r correction on detection. Methods: A receiver operating characteristic (RO C) study was devised to investigate the relative value of a scatter-subtrac tion strategy in SPECT imaging. The study used simulated images of the biod istribution of Tc-99m-labeled FO23C5 anticarcinoembryonic antigen antibodie s within the liver. Projection data for 3 separate tumor locations and 5 st rategies for handling scatter were obtained using Monte Carlo software appl ied to an anthropomorphic phantom. The strategies were (a) perfect scatter rejection, (b) no scatter correction, (c) no scatter correction under an as sumption of an elevated amount of scatter, (d) an energy-spectrum-based sca tter compensation of the normal-scatter case (b), and (e) similar scatter c ompensation for the elevated-scatter case (c), Image reconstruction approxi mated current clinical procedures at the University of Massachusetts Medica l School. Human performance for each combination of location and strategy w as based on averaging the areas under the ROC curve for 7 individuals, A se t of 15 signal-to-noise ratios (SNRs) was derived from these averages for c omparison with SNRs for CHO models featuring constant-Q and difference-of-g aussian (DOG) filters. Results: The Spearman rank correlation coefficient w as 0.92 (P = 0.000001) when comparing task performances for the average hum an and a constant-Q CHO using 4 square-profile channels. For the DOG versio n of the CHO, comparison with the average human found a coefficient of 0.84 (P = 0.00005). Conclusion: The significant positive correlations found bet ween the rankings of the average human observer and the CHOs for our detect ion task indicate that a channelized model observer could eventually serve as a replacement for human observers. The specific CHO models we have used are best suited to screen for significant differences between strategies be fore a human psychophysical study.