Deep Learning Image Reconstruction to Improve Computed Tomography Image Quality of the Phantom with Standard Liver Density

Objective: This study aimed to compare the quality of reconstructed images by deep learning image reconstruction (DLIR) and adaptive statistical iterative reconstruction-V (ASIR-V) techniques at different scan doses using a phantom with liver density. Methods: The Gammex computed tomography (CT) phantom with a standard liver-density insert (ρ_e^w=1.06) was scanned at six different radiation doses (CTDI_vol): 30, 20, 15, 10, 7.5, and 4.5 mGy). Images obtained at each dose were reconstructed using DLIR and ASIR-V. Image quality was analyzed through the imQuest software. The quality of reconstructed images by DLIR at 4.5 mGy (lowest radiation dose) and ASIR-V at 15 mGy (recommended scan dose) were compared using the Bland–Altman method. Results: Across the six doses, DLIR significantly outperformed ASIR-V in key metrics, such as noise, signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), and detectability index (d' ). Bland–Altman analysis indicated that the quality of reconstructed images by DLIR at 4.5 mGy was significantly better to those by ASIR-V at 15 mGy. The noise level of DLIR images at 4.5 mGy was (17.41±0.32), which is significantly lower than that of ASIR-V at 15 mGy (21.17±0.67). At 4.5 mGy, DLIR SNR, CNR, and d’ were (3.21±0.24), (3.42±0.35), and (8.81±0.63), respectively, which are significantly higher than that of ASIR-V at 15 mGy (2.69±0.14), (2.87±0.11), and (5.61±1.28), respectively. Conclusion: In CT scan of focal liver-density lesions using a phantom, DLIR significantly improved the SNR, CNR, and d’ values and reduced image noise compared to ASIR-V. DLIR was able to achieve better quality image reconstruction at 4.5 mGy than the conventional ASIR-V reconstruction at 15 mGy.