Deep-learning image reconstruction algorithms for CT: A task-based image quality assessment of four CT systems using a phantom.

PURPOSE: The purpose of this study was to assess the performance of iterative reconstruction (IR) and deep-learning image reconstruction (DLR) algorithms developed by four CT vendors in terms of image quality. MATERIALS AND METHODS: Acquisitions were performed on an image quality phantom at three dose levels (1.8, 6 and 11 mGy) using four CT systems (further referred to as G-CT, P-CT, U-CT, and C-CT). For each CT, raw data were reconstructed using the commonly used soft tissue kernel and level for IR and DLR algorithms. Noise power spectrum and task-based transfer function were computed to assess noise magnitude, noise texture (fav) and spatial resolution, respectively. Detectability indexes (d') were computed to model the detection of two abdominal lesions. RESULTS: Compared to IR, noise magnitude reduction with DLR was similar for all dose levels for G-CT (-21.1 ± 1.5 [standard deviation (SD)] %) and P-CT (-48.4 ± 0.1 [SD] %) but more pronounced at 1.8 mGy and decreased as the dose level increased for U-CT and C-CT. Noise texture was greater with DLR than IR at all dose levels for all CT systems, except for U-CT, which gave similar fav values. For both inserts, spatial resolution was better with DLR than with IR for all CT systems, except for the low-contrast insert with C-CT at 1.8 and 6 mGy and P-CT at 1.8 mGy. For both simulated lesions and all dose levels, d' values were greater with DLR than with IR by 77.5 ± 8.7 (SD) % for C-CT, 33.7 ± 5.6 (SD) % for G-CT, 112.7 ± 4.7 (SD) % for P-CT and from 158.3 % to 546.6 % on average for U-CT. CONCLUSION: Compared to IR, DLR algorithms reduce the image noise and improve detectability whilst providing similar or better noise texture and spatial resolution.