|
With many attractive attributes, photon counting detectors with many energy bins are being considered for clinical CT
systems. In practice, a large amount of projection data acquired for multiple energy bins must be transferred in real time
through slip rings and data storage subsystems, causing a bandwidth bottleneck problem. The higher resolution of these
detectors and the need for faster acquisition additionally contribute to this issue. In this work, we introduce a new
approach to lossless compression, specifically for projection data from photon counting detectors, by utilizing the
dependencies in the multi-energy data. The proposed predictor estimates the value of a projection data sample as a
weighted average of its neighboring samples and an approximation from other energy bins, and the prediction residuals
are then encoded. Context modeling using three or four quantized local gradients is also employed to detect edge
characteristics of the data. Using three simulated phantoms including a head phantom, compression of 2.3:1-2.4:1 was
achieved. The proposed predictor using zero, three, and four gradient contexts was compared to JPEG-LS and the ideal
predictor (noiseless projection data). Among our proposed predictors, three-gradient context is preferred with a
compression ratio from Golomb coding 7% higher than JPEG-LS and only 3% lower than the ideal predictor. In
encoder efficiency, the Golomb code with the proposed three-gradient contexts has higher compression than block
floating point. We also propose a lossy compression scheme, which quantizes the prediction residuals with scalar
uniform quantization using quantization boundaries that limit the ratio of quantization error variance to quantum noise
variance. Applying our proposed predictor with three-gradient context, the lossy compression achieved a compression
ratio of 3.3:1 but inserted a 2.1% standard deviation of error compared to that of quantum noise in reconstructed images.
From the initial simulation results, the proposed algorithm shows good control over the bits needed to represent multienergy
projection data.
|
