Comparison of Brain Computed Tomography Attenuation Values Between Deep-Learning and Conventional Reconstruction Methods: A Bias-Free Approach.

OBJECTIVES: To investigate the differences in brain computed tomography (CT) attenuation with deep-learning reconstruction (DLR) versus conventional reconstruction methods using a bias-free approach. METHODS: Unenhanced brain CT scans of 27 participants (11 males and 16 females; mean age: 46.4 y) without brain abnormalities were included. The scans were reconstructed using 4 methods: hybrid iterative reconstruction (Hybrid-IR), DLR_mild, DLR_standard, and DLR_strong. All CT images were normalized to Montreal Neurological Institute coordinates, and CT attenuation values and signal-to-noise ratios (SNRs) were measured using 22 white and 10 gray-matter regions of interest. The edge rise distance (ERD) was measured in the right and left lateral ventricles. RESULTS: The mean attenuation between Hybrid-IR and DLR ROIs showed strong correlations in both white (R=0.953 to 0.964) and gray (R=0.914 to 0.943) matter. The Bland-Altman analysis showed greater attenuation with DLR than with Hybrid-IR, with mean biases of 2.27 to 2.40 in white matter and 1.78 to 1.84 in gray matter. The median SNRs were highest for DLR_strong, followed by DLR_standard, DLR_mild, and Hybrid-IR. The median ERD of DLR_mild [2.86 (pixels)] was the lowest, followed by DLR_standard (3.27), Hybrid-IR (3.27), and DLR_strong (3.55). The Freedman test revealed significant differences based on the reconstruction method (P=0.0080). CONCLUSIONS: The attenuation values for Hybrid-IR and DLR were strongly correlated, with DLR showing higher attenuation.