Aldo Badano PhD, Presenter: Research grant, Koninklijke Philips Electronics NV, Saronno, Italy

While medical images are acquired with bit depths of 12 to 16 at the detector, most imaging systems quantize the data to 8 bits for display. It is generally assumed that this loss of information does not translate into a degraded visual task performance. This assumption relies either on the human visual system not able to perceive more than a finite number of distinct luminance levels, or on the wide luminance range adversely affecting the perception of subtle image features due to severe under-adaptation. In this work, we show with model observers that the effect of quantization on observer performance is significant at least for the task of detecting a low contrast signal in complex mammography-like backgrounds.

Synthetic images were generated using the cluster lumpy background (CLB) technique. We compare results for single- and double-layer CLB structures that have been shown to replicate the different correlation structures in digital mammography images. We report the performance in terms of the area under the ROC curve (AUC) results for a channelized Hotelling observer (CHO) with difference-of-Gaussians channels acting on the floating point (FP) grayscale data and on the quantized data for 8- and 16-bit systems. We also report the efficiency of the observers given the quantized data relative to the observers given the FP data by calculating the ratio of squared signal-to-noise ratios.

When using 1000 random single-layer CLBs to estimate each of the covariance and the mean signal, 1000 random testing image pairs, and signal amplitude of 0.05, we obtained CHO performance estimates of 0.756, 0.754, and 0.731 for FP, 16-, and 8-bit systems. For double-layer CLBs, the CHO performance was 0.894, 0.888, and 0.833, respectively. The corresponding efficiencies for the 16- and 8-bit systems are 0.987 and 0.789, and 0.947 and 0.597, for the single- and double-layer CLBs, respectively.

Our results indicate that the quantization errors for the visual task considered are significant (0.02-0.06 in AUC units) with a maximum efficiency drop of 0.4. Our findings suggest that for the task considered, a quantization over 16-bit integers is almost equivalent in terms of performance to no quantization, while an 8-bit quantization significantly reduces detection performance. It remains to be seen if this degradation in performance can also be measured with human readers.

If our findings are verified with human readers, we could extend our analysis to show the potential benefits of higher bit depth display approaches in diagnostic imaging.

Badano, A, Physics Keynote Speaker: Quantization Errors in Radiology—Initial Model Observer Results.  Radiological Society of North America 2008 Scientific Assembly and Annual Meeting, February 18 - February 20, 2008 ,Chicago IL. http://archive.rsna.org/2008/7002490.html