Fig 1.
Schematic diagram illustrating the concept of the generation model by integrating with [18F]-FP-CIT TAC.
At the top left, a small TAC graph shows the SUV over time in the defined VOIs from the PU and OC for both normal participants and patients with PD. When the OC is used as the reference region, the resulting SUVRs over time are plotted in the bottom right graph, which demonstrates a growing difference between normal and PD groups as time progresses. After the peak uptake in the PU, a series of five frames of early PET imaging (each 2 min in duration) acquired between 30 and 40 min post-injection was used as input to a generator in a cGAN model. The generator was trained to produce delayed images that resembled the real delayed PET image acquired at 180 min post-injection. A discriminator simultaneously evaluated whether the generated delayed image could be distinguished from the real delayed image, allowing the generator to improve its output through adversarial training. Abbreviations: TAC, time-activity curve; SUV, specific uptake value; VOI, volumes of interest; PU, putamen; OC, occipital cortex; PD, Parkinson’s disease; SUVR, standardized uptake value ratio; PET, positron emission tomography; cGAN, conditional generative adversarial network.
Table 1.
Patient characteristics of datasets used for the training, internal, and independent validation of the cGAN framework.
Fig 2.
Two representative cases with transaxial and MIP images from paired early-phase, generated delayed-phase, and real delayed-phase PET scans in the internal validation set.
(a–f): Case 1: 73-year-old male with essential tremors showing no significant reduction in uptake in the bilateral striatum. (g–l): Case 2: 40-year-old male with PD showing decreased DAT binding in the striatum, especially the PP. In each case, the transaxial and MIP images are arranged from left to right in the following order: early-phase PET (a, d, g, j), generated delayed-phase PET (b, e, h, k), and real delayed-phase PET (c, f, i, l). Abbreviations: MIP, maximum intensity projection; PET, positron emission tomography; PD, Parkinson’s disease; DAT, dopamine transporter; PP, posterior putamen.
Fig 3.
Three representative cases with transaxial and MIP images from paired early-phase, generated delayed-phase, and real delayed-phase PET scans in the independent validation set.
(a–f): Case 1: 41-year-old male (normal control) demonstrating no significant reduction in uptake in the bilateral striatum. However, the CA in the generated delayed image (b, e) exhibits a relatively shorter contour than the real delayed image. (g–l): Case 2: 59-year-old male with PD revealing decreased DAT binding in the striatum, particularly in the PP, with a ventrodorsal gradient. (m–r): Case 3: 56-year-old male (normal control) with discrepancies observed, particularly in the left CA. Visual analysis classified the generated image (n, q) as non-DP by consensus, while the real image (o, r) was correctly diagnosed with a normal DAT binding. In each case, the transaxial and MIP images are arranged from left to right in the following order: early-phase PET (a, d, g, j, m, p), generated delayed-phase PET (b, e, h, k, n, q), and real delayed-phase PET (c, f, i, l, o, r). Abbreviations: MIP, maximum intensity projection; PET, positron emission tomography; CA, caudate nucleus; PD, Parkinson’s disease; DAT, dopamine transporter; PP, posterior putamen; DP, degenerative parkinsonism.
Fig 4.
Scatter plots showing the correlation between SNBR obtained from generated and real delayed-phase PET images in the internal validation set.
The SNBR of the whole striatum exhibited a high correlation, with an R-value of 0.93. Subregional SNBRs showed high correlations, ranging from 0.90 for the VS to 0.95 for the PP. Abbreviations: SNBR, specific to non-specific binding ratio; PET, positron emission tomography; VS, ventral striatum.
Fig 5.
Scatter plots showing the correlation between SNBR from generated and real delayed-phase PET images in the independent validation set.
The SNBR of the whole striatum exhibited a high correlation, with an R-value of 0.90. Subregional SNBRs exhibited high correlations, ranging from 0.84 for the CA to 0.92 for the PP. Abbreviations: SNBR, specific to non-specific binding ratio; PET, positron emission tomography; PP, posterior putamen. CA, caudate nucleus.
Table 2.
ROC-AUC with permutation tests regarding the diagnosis of DP using subregional SNBRs and ISRs in the internal and independent validation sets.
Table 3.
Diagnostic performance in detecting abnormalities or DP based on consensus interpretations of generated and real delayed-phase imaging in the internal validation set.
Table 4.
Diagnostic performance in detecting abnormalities or DP based on consensus interpretations of generated and real delayed-phase imaging in the independent validation set.
Table 5.
Intra-reader agreement between generated and real delayed-phase images for detecting abnormalities or DP in the internal and independent validation sets.