Effect of Scan Time on Neuro 18F-Fluorodeoxyglucose Positron Emission Tomography Image Generated Using Deep Learning
The purpose of this study was to generate the PET images with high signal-to-noise ratio (SNR) acquired for typical scan durations (H-PET) from short scan time PET images with low SNR (L-PET) using deep learning and to evaluate the effect of scan time on the quality of predicted PET
image. A convolutional neural network (CNN) with a concatenated connection and residual learning framework was implemented. PET data from 27 patients were acquired for 900 s, starting 60 minutes after the intravenous administration of FDG using a commercial PET/CT scanner. To investigate the
effect of scan time on the quality of the predicted H-PETs, 10 s, 30 s, 60 s, and 120 s PET data were generated by sorting the 900 s LMF data into the LMF data acquired for each scan time. Twenty-three of the 27 patient images were used for training of the proposed CNN and the remaining four
patient images were used for test of the CNN. The predicted H-PETs generated by the CNN were compared to ground-truth H-PETs, L-PETs, and filtered L-PETs processed with four commonly used denoising algorithms. The peak signal-to-noise ratios (PSNRs), normalized root mean square errors (NRMSEs),
and average regionof- interest (ROI) differences as a function of scan time were calculated. The quality of the predicted H-PETs generated by the CNN was superior to that of the L-PETs and filtered L-PETs. Lower NRMSEs and higher PSNRs were also obtained from predicted H-PETs compared to the
L-PETs and filtered L-PETs. ROI differences in the predicted H-PETs were smaller than those of the L-PETs. The quality of the predicted H-PETs gradually improved with increasing scan times. The lowest NRMSEs, highest PSNRs, and smallest ROI differences were obtained using the predicted H-PETs
for 120 s. Various performance test results for the proposed CNN indicate that it is possible to generate H-PETs from neuro FDG L-PETs using the proposed CNN method, which might allow reductions in both scan time and injection dose.
Keywords: CNN; Deep Learning; Denoising; Human Brain; PET; Scan Time Reduction
Document Type: Research Article
Affiliations: 1: Department of Electronic Engineering, Sogang University, 35 Baekbeom-ro, Mapo-gu, Seoul 04107, Korea 2: Departments of Nuclear Medicine, Severance Hospital, Yonsei University, 50-1 Yonsei-ro, Seodaemun-gu, Seoul 03722, Korea
Publication date: 01 March 2021
- Journal of Medical Imaging and Health Informatics (JMIHI) is a medium to disseminate novel experimental and theoretical research results in the field of biomedicine, biology, clinical, rehabilitation engineering, medical image processing, bio-computing, D2H2, and other health related areas.
- Editorial Board
- Information for Authors
- Subscribe to this Title
- Ingenta Connect is not responsible for the content or availability of external websites
- Access Key
- Free content
- Partial Free content
- New content
- Open access content
- Partial Open access content
- Subscribed content
- Partial Subscribed content
- Free trial content