Image Quality Assessment for Deep Learning Image Reconstruction Algorithm: A Phantom Study

Objective: To compare the image quality between the deep learning image reconstruction (DLIR) and iterative reconstruction (IR) algorithms via the dedicated phantom. Method: ACR quality assurance phantom (Gammex 464) was scanned by GE Revolution Apex. The CT value accuracy, low contrast resolution, image uniformity and high contrast resolution of five substances from module 1 to module 4 were measured respectively. Through the above indicators, the image quality of three levels (DL, DM and DH) of Truefedelitytm (TFI) and three levels (30%, 60% and 90%) of adaptive statistical iterative reconstruction-V (asir-V, hereinafter referred to as AV) under high dose (20 mgy) were compared. The comparison between the two algorithms for each parameter was tested by One-Way Anova. Results: The high/low-contrast resolution of the six image series were consistent (high-contrast resolution: 10 lp/cm; low-contrast resolution: 6 mm). The two algorithms both slightly overestimated the CT value of polyethylene, air and acrylic, and no statistically significant difference was found among the difference of CT values of the substances. Both algorithms underestimated the CT value of bone and the solid water; TFI showed better performance in evaluating the solid water which was closer to the real value, though statistical difference was not found between each group of images. Among the 6 groups of images, TFI DH showed the best image uniformity and at the same reconstruction level, TFI showed better uniformity than AV. Conclusion: DLIR can further reduced image noise while maintaining image spatial resolution and CT value accuracy at high dose level.