Clinical study of deep learning reconstruction to improve the quality of rapidly acquired PET images
10.3760/cma.j.cn321828-20210514-00164
- VernacularTitle:深度学习重建方法改善快速采集PET图像质量的临床研究
- Author:
Linjun HU
1
;
Yiyi HU
;
Binwei GUO
;
Meng LIANG
;
Xinzhong HAO
;
Zhixing QIN
;
Sijin LI
;
Zhifang WU
Author Information
1. 山西医科大学第一医院核医学科、分子影像精准诊疗省部共建协同创新中心，太原 030001
- Keywords:
Deep learning;
Image processing, computer-assisted;
Positron-emission tomography;
Deoxyglucose
- From:
Chinese Journal of Nuclear Medicine and Molecular Imaging
2021;41(10):602-606
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To improve the quality of 18F-fluorodeoxyglucose (FDG) PET images at different acquisition times through deep learning (DL) PET image reconstruction methods. Methods:A total of 45 patients (20 males, 25 females; age (52.0±13.6) years) with malignant tumors and PET/CT scans from September 2020 to October 2020 in the Department of Nuclear Medicine of the First Hospital of Shanxi Medical University were included in this retrospective study. The short acquisition time 30 s/bed PET images from the raw list mode were selected as the input of DL model. DL image reconstruction model, based on the Unet algorithm, was trained to output imitated PET images with full dose standard acquisition time (3 min). The image quality evaluation and quantitative analysis were carried out for four groups of images: DL images, 30 s, 90 s, and 120 s images, respectively. The quality of PET images in four groups was evaluated using the five-point method. Liver background activities, lesions quantification parameters (maximum standardized uptake value (SUV max), mean standardized uptake value (SUV mean), standard deviation (SD), signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR)), and first-order texture features (skewness, kurtosis, uniformity, entropy) were measured. Kappa test, χ2 test and one-way analysis of variance (least significant difference t test) were used for data analysis. Results:The image quality scores between four groups were highly consistent ( Kappa=0.799, P<0.001). The number of patients with scores≥3 in DL, 30 s, 90 s and 120 s groups were 6, 4, 7 and 8, respectively ( χ2=125.47, P<0.001). The liver SD of DL group was significantly lower than that of 30 s group (0.26±0.07 vs 0.43±0.11; F=3.58, t=-7.91, P<0.05). The liver SNR of DL group was higher than that of 30 s group (11.04±4.36 vs 5.41±1.41; F=10.22, t=5.40, P<0.05). The liver SD and SNR of DL group were similar to those of 90 s group (0.39±0.16, 8.46±3.34; t values: -0.87 and 2.17, both P>0.05). In 18 tumor lesions with high uptake, SNR and CNR of DL group were significantly higher than those of 30 s group (60.21±29.26 vs 38.38±16.54, 22.26±15.85 vs 15.41±9.51; F values: 13.09 and 7.05; t values: 5.20 and 4.04, both P<0.001). There were statistically significant differences among four groups in the first-order texture features ( F values: 4.30-9.65, all P<0.05), but there was no significant difference between DL group and 120 s group ( t values: from -1.25 to 0.15, all P>0.05). Conclusion:DL reconstruction model can improve the quality of short-frame PET images, which meets the needs of clinical diagnosis, efficacy evaluation and radiomics research.
