Soft-tissue detectability in cone-beam computed tomography (CBCT) was evaluated via two-alternative forced-choice (2AFC) tests. Investigations included the dependence of detectability on radiation dose, the influence of the asymmetric three-dimensional (3D) noise-power spectrum (NPS) in axial and sagittal or coronal planes, and the effect of prior knowledge on detectability. Custom-built phantoms (approximately 15 cm diameter cylinders) containing soft-tissue-simulating spheres of variable contrast and diameter were imaged on an experimental CBCT bench. The proportion of correct responses (Pcorr) in 2AFC tests was analyzed as a figure of merit, ideally equal to the area under the receiver operating characteristic curve. Pcorr was evaluated as a function of the sphere diameter (1.6-12.7 mm), contrast (20-165 HU), dose (1-7 mGy), plane of visualization (axial/sagittal), apodization filter (Hanning and Ram-Lak), and prior knowledge provided to the observer [ranging from stimulus known exactly (SKE) to stimulus unknown (SUK)]. Detectability limits were characterized in terms of the dose required to achieve a given level of Pcorr (e.g., 70%). For example, a 20 HU stimulus of diameter down to approximately 6 mm was detected with Pcorr 70% at dose > or =2 mGy. Detectability tended to be greater in axial than in sagittal planes, an effect amplified by sharper apodization filters in a manner consistent with 3D NPS asymmetry. Prior knowledge had a marked influence on detectability--e.g., Pcorr for a approximately 6 mm (20 HU) sphere was approximately 55%-65% under SUK conditions, compared to approximately 70%-85% for SKE conditions. Human observer tests suggest practical implications for implementation of CBCT: (i) Detectability limits help to define minimum-dose imaging techniques for specific imaging tasks; (ii) detectability of a given structure can vary between axial and sagittal/coronal planes, owing to the spatial-frequency content of the 3D NPS in relation to the imaging task; and (iii) performance under SKE conditions (e.g., image guidance tasks in which lesion characteristics are known) is maintained at a lower dose than in SUK conditions (e.g., diagnostic tasks in which lesion characteristics are unknown).