Mitsunori Goto MMedSc, RT, Presenter: Nothing to Disclose

Masaaki Taura BMedSc, RT, Abstract Co-Author: Nothing to Disclose

Kazuhiro Sato MMedSc, RT, Abstract Co-Author: Nothing to Disclose

Noriyasu Homma PhD, Abstract Co-Author: Nothing to Disclose

Issei Mori, Abstract Co-Author: Nothing to Disclose

ROI must be reasonably large for low frequency accuracy even with proper windowing. If ROI needs to be very small, deconvolution is a choice.  

To evaluate noise reduction performance of iterative reconstruction (IR), noise standard deviation is not a good noise indicator and noise power spectrum (NPS) analysis is needed. For the evaluation of low-contrast detection performance, low-frequency component of NPS is crucially important because signal exists only at low frequency region. On the other hands, NPS measurement of CT image is inaccurate at low frequency due to frequency leakage problem. We show the low-frequency error of NPS quantitatively in association with the size of region of interest (ROI) and usage of windowing. We further show that the frequency leakage can be corrected by a deconvolution.  

Overestimate of low frequency NPS worsens with smaller ROI. The frequency leakage problem is dominated by the length of short side of rectangular ROI. Windowing is effective to suppress this error, but becomes almost powerless if ROI is 32x32 pixels or smaller. Among several window functions, we judged Welch type is the most preferable. When ROI is 32x32 pixels, MFSNR for 10 or 20mm object size is underestimated by a factor of more than 10% even with windowing. This error can be made virtually zero by deconvolution.

We generated 512 independent noise images by PC simulation. Their true NPS is theoretically known and statistically identical with that of clinical 512x512 matrix images of 320 mm FOV using Shepp-Logan kernel. Rectangular ROIs of various sizes are set, and 2-dimensional NPS of each ROI is obtained. All 2D-NPSs are transformed to 1-D NPS by circumferential averaging. The average NPS is obtained from 512 NPSs for each ROI size. Windowing is performed by multiplying a window function to pixel values within ROI before applying to 2D-FT. A Richardson-Lucy type deconvolution is applied to 2-D NPS. Deconvolved 2-D NPS is then converted to 1-D NPS. For the deconvolution, the kernel was designed such that it approximates macroscopic frequency leakage effect. The matched filter SNR (MFSNR) was calculated for each of non-windowed, windowed, and deconvolved NPS.  

Goto, M, Taura, M, Sato, K, Homma, N, Mori, I, Measurement of Noise Power Spectrum for CT Image: Importance of Low Frequency Component and Methods to Achieve Its Accuracy.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14012829.html