Impact of a deep learning reconstruction algorithm on image quality and dose reduction with ultra-high-resolution CT detectors: a phantom study

2026-02-12

Yuhe Cheng, Zixuan Ma, Senlin Guo, Chensi Xu, Dandan Liu, Yongxian Zhang,
Impact of a deep learning reconstruction algorithm on image quality and dose reduction with ultra-high-resolution CT detectors: a phantom study,
Zeitschrift für Medizinische Physik,
2025,
,
ISSN 0939-3889,
https://doi.org/10.1016/j.zemedi.2025.11.002.
(https://www.sciencedirect.com/science/article/pii/S0939388925001527)
Abstract: Purpose
To quantitatively evaluate the image quality and the radiation dose reduction potential when a deep learning reconstruction (DLR) algorithm is combined with an ultra-high-resolution (UHR) detector, using a task-based assessment framework (MTF, NPS, TTF, and detectability index d′).
Methods
A Catphan 600 phantom was scanned at five CTDIvol levels (CTDIvol: 10, 7.5, 5, 2.5, 1 mGy). Data acquired with collimation width of 64 × 0.625 mm were reconstructed with Filtered Back-Projection (FBP) and adaptive statistical iterative reconstruction (ClearView 50 %, CV50); for 128 × 0.3125 mm, five algorithms were applied: FBP, CV50, and ClearInfinity (deep learning reconstruction algorithm) 10 % (CI10), 50 % (CI50), 90 % (CI90). The Modulation Transfer Function (MTF), Noise Power Spectrum (NPS), Task Transfer Function (TTF), and Detectability Index (d′) for large, subtle and small features were measured.
Results
At all dose levels, MTF50% and MTF10% with a 0.3125 mm collimation width was higher than that with 0.615 mm, improving the d′ of small features but increasing noise. CI markedly reduced NPS peaks without shifting average spatial frequency, thereby increasing d′ for large and subtle features. The combination of the two achieved the lowest noise peak and the highest detectability index.
Conclusions
Integrating a deep learning reconstruction algorithm with a UHR detector enhances spatial resolution, reduces noise magnitude without texture alteration, improves lesion detectability, and demonstrates substantial potential for radiation dose reduction.
Keywords: Computed tomography; Deep learning; Image reconstruction; Image enhancement; Ultra-high-resolution computed tomography