Feasibility Study of Deep Learning-Driven Image Restoration for Fast, Low-Count [<sup>18</sup>F]FP-CIT Digital PET/CT.
Authors
Affiliations (5)
Affiliations (5)
- Department of Nuclear Medicine, Kangdong Sacred Heart Hospital, Hallym University College of Medicine, 150, Seongan-ro, Gangdong-gu, Seoul 05355, Republic of Korea.
- Department of Nuclear Medicine, Hanyang University Medical Center, Hanyang University College of Medicine, 221-1, Wangsimni-ro, Seongdong-gu, Seoul 04763, Republic of Korea.
- Department of Neurology, Uijeongbu Eulji Medical Center, Eulji University School of Medicine, 712, Dongil-ro, Uijeongbu-si 11759, Republic of Korea.
- Department of Nuclear Medicine, Uijeongbu Eulji Medical Center, Eulji University School of Medicine, 712, Dongil-ro, Uijeongbu-si 11759, Republic of Korea.
- Department of Radiological Science, College of Health Science, Eulji University, 553, Sanseong-daero, Sujeong-gu, Seongnam-si 13135, Republic of Korea.
Abstract
<b>Background/Objectives</b>: N-3-[<sup>18</sup>F]fluoropropyl-2β-carbomethoxy-3β-4-iodophenyl nortropane ([<sup>18</sup>F]FP-CIT) positron emission tomography/computed tomography (PET/CT) is an effective imaging tool for diagnosing parkinsonism. This study evaluated the feasibility of deep learning (DL)-driven image restoration of short-duration [<sup>18</sup>F]FP-CIT images. <b>Methods</b>: List-mode data from 202 patients who underwent [<sup>18</sup>F]FP-CIT PET/CT were reconstructed into 30-s (30<sub>sec</sub>), 1-min (1<sub>min</sub>), and reference 10-min (10<sub>min</sub>) acquisition durations. Patients were divided into training, validation, and test sets in a 6:2:2 ratio. The U-Net model was implemented to generate the DL images from 30<sub>sec</sub> and 1<sub>min</sub> data. The visual image quality was assessed using a three-point scale, visual interpretation of striatal dopamine transporter binding patterns, regional standardized uptake value ratios (SUVRs), and quantitative quality metrics including the peak signal-to-noise ratio, root mean squared error, and universal quality index among five series of images: 30<sub>sec</sub>, 1<sub>min</sub>, DL-30<sub>sec</sub>, DL-1<sub>min</sub>, and 10<sub>min</sub>. <b>Results</b>: While 30<sub>sec</sub> and 1<sub>min</sub> low-count scans showed poor image quality, the DL-driven algorithm showed significant improvement; DL-1<sub>min</sub> scans achieved excellent ratings in 95% of cases. Concordance with the reference images was 85.0% for visual image quality and 92.5% for visual interpretation in the DL-30<sub>sec</sub> images, and 97.5% and 92.5%, respectively, in the DL-1<sub>min</sub> images. Discordant interpretations occurred mainly in patients with atypical parkinsonism. Regional SUVR values and quantitative metrics for the DL-1<sub>min</sub> images showed good agreement and smaller biases with respect to the reference images, compared with those of DL-30<sub>sec</sub> images. <b>Conclusions</b>: The DL-driven method generated clinically acceptable [<sup>18</sup>F]FP-CIT images while reducing acquisition time, with better image quality in the DL-1<sub>min</sub> than in the DL-30<sub>sec</sub> images. Concordance in visual interpretation was reduced in the small subgroup of patients with atypical parkinsonism.