MENU

Click here to

Ă—

Are you sure ?

Yes, do it No, cancel

Quantitative Comparison of Noise Texture in Gemstone Spectrum Imaging CT Images Reconstructed Using Filtered Back-Projection (FBP), Iterative Reconstruction, and Deep Learning Techniques

J Tang*, B Nett, P Prakash, GE Healthcare Technologies, Waukesha, WI

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: quantitatively compare noise texture via noise power spectra of computed tomography (CT) images reconstructed using filtered back-projection (FBP), iterative reconstruction (ASiR-V), and deep learning image reconstruction (DLIR) in Gemstone Spectrum Imaging.

Methods: assess the noise texture across image reconstruction algorithms, we calculated the normalized noise power spectra (nNPS) of uniform phantom images scanned with dual energy Gemstone Spectrum Imaging (GSI), and reconstructed using FBP, iterative reconstruction (ASiR-V, 100%), and deep learning (DLIR, high).
A 30cm water phantom was scanned on Revolution CT (GE Healthcare, Waukesha, WI) (40mm axial, 80/140kV switching, 475mA, 1s rotation, CTDIvol 23.59mGy), and 0.625mm thick 40 keV, 70 keV, 140 keV, water/iodine and iodine/water images were reconstructed using the three algorithms above. Normalized NPS (nNPS) was computed for each type of GSI image.
To obtain the nNPS, first, the difference image between adjacent slices was calculated to get noise only images for the center 32 slice locations. Then, the 2D NPS of the noise only image was calculated, normalized to its own area, and radially averaged to yield the normalized NPS (nNPS). Finally the nNPS from all 32 slice locations were averaged to get the nNPS for comparison. To compare the nNPS, the root-mean squared of nNPS difference (RMSD) between nNPSASiR-V /nNPSDLIR and the corresponding nNPSFBP was calculated.

Results: between nNPSDLIR and nNPSFBP were 0.14, 0.18, 0.30, 0.38, 0.21 for 40 keV, 70 keV, 140 keV, water and iodine images respectively; while the corresponding RMSDs between nNPSDLIR and nNPSFBP were 1.28, 1.26, 1.30, 1.41, 1.27. The nNPS of images reconstructed with DLIR and FBP showed much closer match compared to ASiR-V.

Conclusion: with previous reports, normalized NPS of ASiR-V images is shifted towards lower spatial frequencies. The normalized NPS of DLIR closely matches FBP images across various types of images offered by GSI imaging as quantified via RMSD.

Funding Support, Disclosures, and Conflict of Interest: Employees of GE Healthcare

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email