Comparison of noise-power spectrum and modulation-transfer function for CT images reconstructed with iterative and deep learning image reconstructions: An initial experience study

Abstract Introduction Deep learning image reconstruction (DLIR) is a very recent image reconstruction method that is already available for commercial use. We evaluated the quality of DLIR images and compared it to the quality of images from the latest adaptive statistical iterative reconstruction (ASIR-V) algorithm in terms of noise-power spectrum (NPS) and modulation-transfer function (MTF). Methods We scanned a Revolution QA phantom (GE Healthcare, USA) and a 20 cm water phantom (GE Healthcare, USA) with our 512 multi-slice computed tomography (CT) scanner. Images of the tungsten wire within the Revolution QA phantom were reconstructed with a 50 mm field of view (FOV). The images were reconstructed with various ASIR-V strengths (i.e. 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100%) and DLIRs (i.e. low, medium, and high) to assess the MTF. The images from the 20 cm water phantom were reconstructed with the same configuration to assess the NPS. Results The MTF was similar for both reconstruction algorithms of DLIR and ASiR-V. The peak frequency (fp) of the DLIR low was comparable to that from ASIR-V at 50, 60, 70%; the DLIR medium was comparable to ASIR-V at 80%; and the DLIR high was comparable to ASIR-V at 100%. The average frequency (fA) of the DLIR low was comparable to that from ASIR-V at 40%; the DLIR medium was comparable to ASIR-V at 50%; and the DLIR high was comparable to ASIR-V at 70%. Both the DLIR and ASIR-V were able to reduce noise, but they had a different texture. Conclusions The noise in the DLIR images was more homogenous at high and low frequencies, while in the ASIR-V images, the noise was more concentrated at high frequencies. The MTF was similar for both reconstruction algorithms. The DLIR method showed a better noise reduction than the ASIR-V reconstruction.