Assessment of prostate cancer detection with a visual-search observer in SPECT-CT imaging

Objectives The potential of SPECT for staging prostate cancer has been slow to develop. Current protocols with In-111 Prostascint use medium-energy, parallel-hole (MEPH) geometries. As a preliminary step towards examining alternative geometries, we compared the suitability of several existing mathematical models as human-model observers for assessing tumor detection and localization in reconstructed images.Methods An analytic projector imaged a family of XCAT phantoms with clinically realistic Prostascint uptake in the pelvic and lower abdomen regions. Spherical tumors (1-cm diameter) were placed in the prostate and pelvic lymph nodes. The projector modeled the attenuation of 171 and 245 keV photons and distance-dependent MEPH camera blur for 120 projections over 360 degrees. Count levels and other acquisition parameters were based on clinical studies. Reconstructed volumes were obtained with 2 and 6 iterations (15 angles per subset) of the OSEM algorithm. Gaussian post-filtering with integral FWHMs from 0 through 4 voxels was also applied. Pilot LROC studies were conducted with 2D image slices for a scanning channelized nonprewhitening (CNPW) observer, a visual-search (VS) model observer, and a group of human observers. Each observer read 150 2D images per pairing of iteration and blur level. Anatomical information about node and prostate location in each slice was also provided to all observers.Results Table 1 summarizes detection accuracy for the various observers as quantified by area under the LROC curve. The reconstruction parameters had little effect on the performance of the CNPW observer. The VS and human observers demonstrated good correlation (coefficient 0.82).Conclusions The SPECT-CT study format is preferable to a SPECT format for comparing human observers with these model observers. Implicit modeling of human-observer uncertainty with a VS observer leads to an improved human-observer model compared to the CNPW observer.