Deep learning reconstruction for low‐dose thoracic CT of calcification inserts: A phantom study of CNR, noise magnitude, and noise texture
Journal of Applied Clinical Medical Physics · 2026-12 · 27 (12)
DOI: 10.1002/acm2.70916 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Applied Clinical Medical Physics, online erschienen am 20.11.2026. Verfasst von Ya‐Yun Hsiao und 3 weiteren Autor:innen. Die Arbeit nennt 11 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Purpose
To compare filtered back projection (FBP), 50% Adaptive Statistical Iterative Reconstruction‐Veo (ASiR), and medium‐strength deep learning image reconstruction (DLIR) for hydroxyapatite calcification inserts under different dose, tube‐voltage, and phantom‐size conditions using contrast‐to‐noise ratio (CNR), background noise magnitude, and noise power spectrum (NPS) analysis.
Methods
A QRM‐Thorax containing a QRM‐CCI D100 cardiac calcification insert was scanned on a GE Revolution Apex CT system using institution‐defined routine and low‐dose protocols at 80, 100, 120, and 140 kVp, with and without an Extension Ring L. Images were reconstructed at 2.5‐mm thickness using FBP, ASiR, and DLIR. CNR was evaluated for 3‐ and 5‐mm inserts at 200, 400, and 800 mg HA/cm 3 . Background noise was assessed from a calcification‐excluded mask. For NPS analysis, 10 calcification‐free images were analyzed for each acquisition condition. Total two‐dimensional NPS power, power‐weighted normalized radial NPS, and power‐weighted mean spatial frequency were calculated.
Results
DLIR reduced background noise magnitude and increased CNR relative to FBP across the evaluated conditions, whereas differences relative to ASiR varied with insert size, density, tube voltage, and phantom size. Across all 16 protocol, phantom‐size, and tube‐voltage combinations, ASiR and DLIR reduced total 2D NPS power by 58.1% and 75.8%, respectively, relative to FBP; DLIR reduced total power by 41.9% relative to ASiR. ASiR produced a more consistent shift toward lower power‐weighted mean spatial frequencies, averaging 12.4% below FBP, whereas the shift with DLIR was smaller and more condition dependent, averaging 3.2% below FBP.
Conclusions
In this controlled phantom study, DLIR reduced total noise power and improved CNR relative to FBP, but its incremental difference from ASiR was task and protocol dependent. CNR and noise magnitude should be interpreted together with NPS because nonlinear reconstruction methods can alter both noise magnitude and spatial‐frequency distribution.
Zitieren
APA
Hsiao, Y., Lu, C., Kao, H., & Chen, T. (2026). Deep learning reconstruction for low‐dose thoracic CT of calcification inserts: A phantom study of CNR, noise magnitude, and noise texture. Journal of Applied Clinical Medical Physics, 27(12). https://doi.org/10.1002/acm2.70916
Vancouver
Hsiao Y, Lu C, Kao H, Chen T. Deep learning reconstruction for low‐dose thoracic CT of calcification inserts: A phantom study of CNR, noise magnitude, and noise texture. Journal of Applied Clinical Medical Physics. 2026;27(12). doi:10.1002/acm2.70916
BibTeX
@article{Hsiao2026,
author = {Hsiao, Ya‐Yun and Lu, Cheng‐Chang and Kao, Hui‐Chuan and Chen, Tou‐Rong},
title = {Deep learning reconstruction for low‐dose thoracic CT of calcification inserts: A phantom study of CNR, noise magnitude, and noise texture},
journal = {Journal of Applied Clinical Medical Physics},
year = {2026},
volume = {27},
number = {12},
doi = {10.1002/acm2.70916}
}
EUVIMED