Optimization of estimator performance and comparison to human classification performance as applied to thoracic Ga-67 SPECT images

Quantitative SPECT is believed to have the potential to both improve and expand upon the current diagnostic capabilities of SPECT. The benefits of quantitation will depend, however, on the accuracy and reliability of the estimator. The authors have previously evaluated the choice of estimator for the task of determining the activity of a detected tumor within a transverse slice of a thoracic Ga-67 SPECT scan. Here, the authors examine estimator performance for this same task over a range of reconstruction protocols. Reconstruction is by filtered backprojection with Chang attenuation correction and 2D or 3D Butterworth postfiltering using filter cut-offs from 0.025 cycles/pixel to no cut-off. Two estimators are considered which differ in that the region of interest (ROI) boundaries are determined from the original tumor geometry in one and by edge-detection in the other. Estimator performance is evaluated based on a measure which equally weights bias squared and variance. Calculation of the measure is based upon 35 clinically relevant locations and 200 noise realizations per location. Comparison is made to human-observer lesion-detection performance as evaluated with a localization ROC experiment with similar images. Results indicate that there is an optimal choice of filter cut-off for the estimation task which differs from that for classification. 2D and 3D post-filtering result in equivalent optimal estimator performance, contrary to what has been seen for classification performance. In conclusion, optimal estimation depends upon the reconstruction protocol and the dependence is markedly different than that for classification.