Factors affecting low-contrast feature detectability in computed tomography have been investigated using signal detection experiments. Detectability of a circular object in a uniform noise field was found to increase with decreasing window width. With increasing viewing distance, detectability increased to a maximum before decreasing; the optimal viewing distance was found to increase with decreasing window width. Detectability was independent of window level and image luminance, over the range of values studied. A mathematical model of the CT display system and the human visual system was constructed, using empirical transfer characteristics as system descriptors and an optimal matched filter model of pattern recognition. Theoretical signal-to-noise ratios were in qualitative agreement with the experimental data when account was taken of the visual contrast sensitivity function and the effect of noise internal to the observer. Quantitative discrepancies suggest either that visual matched filter processing is sub-optimal, or that use of the contrast sensitivity function as a spatial filter is only approximately valid. Compared with the visual system characteristics, those of the display system were insignificant. Modifications to CT display configurations are suggested in order to maximise information transfer between the display system and the visual system of the radiologist.