Effects of image quantization and regularization on lesion detectability in SPECT MAP reconstruction

In penalized likelihood methods for SPECT, the degree of smoothing β controls a noise-resolution tradeoff and would seem to be a crucial parameter in controlling image quality as measured by the SNR of a channelized Hotelling observer (CHO) in an SKE/BKE lesion detection task. Application of the CHO to a floating point reconstruction shows a very weak dependence of SNK and β, however, and this dependence is only slightly changed if we modify the floating point reconstruction to the 8-bit quantized version that would be viewed by a human observer. SKE/BKE human studies show a fairly different behavior than that of the CHO, resulting in a strong dependence of SNR on β. Attempts to account for the discrepancy by modeling quantization noise in the CHO fail. We repeat the human and model observer studies at more extreme levels of quantization and note some interesting effects.