Deep learning approximation of attenuation maps for myocardial perfusion SPECT with an IQ·\documentclass[12pt]{minimal}\usepackage{amsmath}\usepackage{wasysym}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{amsbsy}\usepackage{mathrsfs}\usepackage{upgreek}\setlength{\oddsidemargin}{-69pt}\begin{document}$$\varvec{\cdot {}}$$\end{document}SPECT collimator

EJNMMI Physics

Original Research

Reinhard Zabel¹  Tamino Huxohl²  Gopesh Patel²  Wolfgang Burchert² 

[1] Institute of Nuclear Medicine, Hospital Lippe, Lippe, Germany;Institute of Radiology, Nuclear Medicine and Molecular Imaging, Heart and Diabetes Center North Rhine-Westphalia, University Hospital of the Ruhr University Bochum, Bad Oeynhausen, Germany;

关键词: Deep learning;    SPECT;    Myocardial perfusion imaging;    Attenuation correction;   

DOI  :  10.1186/s40658-023-00568-1

 received in 2023-04-05, accepted in 2023-08-07,  发布年份 2023

来源: Springer

【 摘 要 】

BackgroundThe use of CT images for attenuation correction of myocardial perfusion imaging with single photon emission computer tomography (SPECT) increases diagnostic confidence. However, acquiring a CT image registered to a SPECT image is often not possible because most scanners are SPECT-only. It is possible to approximate attenuation maps using deep learning, but this has not yet been shown to work for a SPECT scanner with an IQ·\documentclass[12pt]{minimal}\usepackage{amsmath}\usepackage{wasysym}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{amsbsy}\usepackage{mathrsfs}\usepackage{upgreek}\setlength{\oddsidemargin}{-69pt}\begin{document}$$\varvec{\cdot {}}$$\end{document}SPECT collimator. This study investigates whether it is possible to approximate attenuation maps from non-attenuation-corrected (nAC) reconstructions acquired with a SPECT scanner equipped with an IQ·\documentclass[12pt]{minimal}\usepackage{amsmath}\usepackage{wasysym}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{amsbsy}\usepackage{mathrsfs}\usepackage{upgreek}\setlength{\oddsidemargin}{-69pt}\begin{document}$$\varvec{\cdot {}}$$\end{document}SPECT collimator.MethodsAttenuation maps and reconstructions were acquired retrospectively for 150 studies. A U–Net was trained to predict attenuation maps from nAC reconstructions using the conditional generative adversarial network framework. Predicted attenuation maps are compared to real attenuation maps using the normalized mean absolute error (NMAE). Attenuation-corrected reconstructions were computed, and the resulting polar maps were compared by pixel and by average perfusion per segment using the absolute percent error (APE). The training and evaluation code is available at https://gitlab.ub.uni-bielefeld.de/thuxohl/mu-map.ResultsPredicted attenuation maps are similar to real attenuation maps, achieving an NMAE of 0.020±0.007. The same is true for polar maps generated from reconstructions with predicted attenuation maps compared to CT-based attenuation maps. Their pixel-wise absolute distance is 3.095±3.199, and the segment-wise APE is 1.155±0.769.ConclusionsIt is feasible to approximate attenuation maps from nAC reconstructions acquired by a scanner with an IQ·\documentclass[12pt]{minimal}\usepackage{amsmath}\usepackage{wasysym}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{amsbsy}\usepackage{mathrsfs}\usepackage{upgreek}\setlength{\oddsidemargin}{-69pt}\begin{document}$$\varvec{\cdot {}}$$\end{document}SPECT collimator using deep learning.

【 授权许可】

CC BY   
© Springer Nature Switzerland AG 2023

【 图 表 】

Fig. 3

Fig. 6

Fig. 2

Fig. 2

Fig. 1

Fig. 1

Fig. 4

12888_2023_5124_Article_IEq2.gif

12888_2023_5124_Article_IEq1.gif

Fig. 2

Fig. 3

Fig. 3

Fig. 7

Fig. 2

Fig. 5

Fig. 3

Fig. 4

Fig. 3

Fig. 2

Fig. 1

Fig. 2

Fig. 6

Fig. 4

Fig. 7

Fig. 6

Fig. 2

Fig. 3

Fig. 2

Fig. 1

Fig. 4

Fig. 5

Fig. 2

Fig. 4

Fig. 5

Fig. 1

12862_2023_2133_Article_IEq80.gif

Fig. 6

Fig. 1

Fig. 1

Fig. 1

Fig. 4

Fig. 1

12862_2023_2133_Article_IEq56.gif

Fig. 2