We studied the effect of image quantization by comparing observer detection performance with 8-and 16-bit grayscale presentation. Eight readers evaluated 532 image pairs using a two-alternative forced choice experimental design. The image set consisted of synthetic backgrounds generated using the mammography-like cluster lumpy background (CLB) technique with a dual-layer approach with parameter values that have been shown to replicate the correlation structure found in digital mammography. The image pairs were reviewed in a display device prototype with one million pixels capable of processing and displaying 16-bit images (up to 65536 luminance values). These image pairs were presented either as non-quantized (full range) images in a 16-bit presentation scale, or as quantized, 8-bit images, with a perceptual mapping of gray levels to luminance. The difference in reader performance between reads on quantized image pairs and reads on non-quantized image pairs were derived using fraction of correct decisions. The variance of our measurements was estimated using a multi-reader, multi-case analysis. Average reader performance difference between 16-and 8-bit quantization was 0.065 with an associated standard deviation of 0.048. Our study showed that image quantization is an important factor in visual detection task, that is, a quantization from 16-to 8-bit significantly reduces reader detection performance.
