Generation of synthetic PET images of synaptic density and amyloid from ¹⁸F-FDG images using deep learning

Abstract

Purpose:

Positron emission tomography (PET) imaging with various tracers is increasingly used in Alzheimer’s disease (AD) studies. However, access to PET scans using new or less-available tracers with sophisticated synthesis and short half-life isotopes may be very limited. Therefore, it is of great significance and interest in AD research to assess the feasibility of generating synthetic PET images of less-available tracers from the PET image of another common tracer, in particular ¹⁸F-FDG.

Methods:

We implemented advanced deep learning methods using the U-Net model to predict ¹¹C-UCB-J PET images of synaptic vesicle protein 2A (SV2A), a surrogate of synaptic density, from ¹⁸F-FDG PET data. Dynamic ¹⁸F-FDG and ¹¹C-UCB-J scans were performed in 21 participants with normal cognition (CN) and 33 participants with Alzheimer’s disease (AD). Cerebellum was used as the reference region for both tracers. For ¹¹C-UCB-J image prediction, four network models were trained and tested, which included 1) ¹⁸F-FDG SUV ratio (SUVR) to ¹¹C-UCB-J SUVR, 2) ¹⁸F-FDG K_i ratio to ¹¹C-UCB-J SUVR, 3) ¹⁸F-FDG SUVR to ¹¹C-UCB-J distribution volume ratio (DVR), and 4) ¹⁸F-FDG K_i ratio to ¹¹C-UCB-J DVR. The normalized root mean square error (NRMSE), structure similarity index (SSIM), and Pearson’s correlation coefficient were calculated for evaluating the overall image prediction accuracy. Mean bias of various ROIs in the brain and correlation plots between predicted images and true images were calculated for ROI-based prediction accuracy. Following a similar training and evaluation strategy, ¹⁸F-FDG SUVR to ¹¹C-PiB SUVR network was also trained and tested for ¹¹C-PiB static image prediction.

Results:

The results showed that all four network models obtained satisfactory ¹¹C-UCB-J static and parametric images. For ¹¹C-UCB-J SUVR prediction, the mean ROI bias was −0.3% ± 7.4% for the AD group and −0.5% ± 7.3% for the CN group with ¹⁸F-FDG SUVR as the input, −0.7% ± 8.1% for the AD group, and −1.3% ± 7.0% for the CN group with ¹⁸F-FDG K_i ratio as the input. For ¹¹C-UCB-J DVR prediction, the mean ROI bias was −1.3% ± 7.5% for the AD group and −2.0% ± 6.9% for the CN group with ¹⁸F-FDG SUVR as the input, −0.7% ± 9.0% for the AD group, and −1.7% ± 7.8% for the CN group with ¹⁸F-FDG K_i ratio as the input. For ¹¹C-PiB SUVR image prediction, which appears to be a more challenging task, the incorporation of additional diagnostic information into the network is needed to control the bias below 5% for most ROIs.

Conclusions:

It is feasible to use 3D U-Net-based methods to generate synthetic ¹¹C-UCB-J PET images from ¹⁸F-FDG images with reasonable prediction accuracy. It is also possible to predict ¹¹C-PiB SUVR images from ¹⁸F-FDG images, though the incorporation of additional non-imaging information is needed.

1 |. INTRODUCTION

Alzheimer’s disease (AD) is an insidious neurodegenerative disorder resulting in significant progressive cognitive and functional decline, and is the most common cause of dementia, accounting for approximately 60%–80% of cases. AD pathology is traditionally characterized by a distinct accumulation of extracellular β-amyloid plaques and intracellular neurofibrillary tangles (NFTs) of tau,^1,2 but also demonstrates significant spatiotemporal alterations in synaptic density and glucose metabolism.³ Accordingly, the clinical symptoms of AD are related to a distinct pathology, including β-amyloid plaques and neurofibrillary tangles, and are well-correlated with synaptic loss and abnormal glucose metabolism in the brain. Positron emission tomography (PET) imaging has been increasingly employed in AD studies to detect this accumulation of β-amyloid⁴ and NFTs and alterations in glucose metabolism⁵ and synaptic density^3,6 with specific tracers in both cross-sectional and longitudinal models.

Traditionally, ¹⁸F-FDG-PET has been used as a surrogate marker of disease progression, measuring synaptic activity via glucose metabolism.⁷ However, glucose metabolism cannot be interpreted as a direct measure of synaptic density, and ¹⁸F-FDG can easily be confounded by medications, sensory stimulation, and glucose levels.^8,9 Alternatively, to assess synaptic loss more directly, one kind of suitable target is the synaptic vesicle glycoprotein 2 (SV2). One of its isoforms, SV2A, is ubiquitously expressed in neurons and is involved in the regulation of neurotransmitter release.⁶ A recently developed tracer, ¹¹C-UCB-J, has been demonstrated to detect SV2A quantitatively in the brain, giving the possibility for detecting synaptic density in the brain directly.¹⁰ Furthermore, recent brain PET imaging studies have demonstrated that SV2A is a promising biomarker to quantify the loss of synaptic density in the AD.^11,12 Besides, β-amyloid plaques and abnormal tau protein deposition in the form of neurofibrillary tangles are also two defining pathological indicators and hallmarks of AD.¹³ Pittsburgh Compound-B (PiB) is a widely used PET tracer to provide quantitative information on fibrillar amyloid deposits.² Among these tracers, ¹¹C-UCB-J and ¹¹C-PiB are labeled with ¹¹C, which has a 20-min half-life. The generation of those tracers requires an onsite cyclotron, which is not easily accessible in most research centers. Besides, it can often be burdensome for patients to have multiple PET scans. Therefore, it is of potential significance and interest in AD research to assess whether synthetic ¹¹C-UCB-J or ¹¹C-PiB images can be conveniently created from a single scan of the most clinically available tracer, ¹⁸F-FDG.

Deep learning has been in a wide variety of fields, including computer vision, speech recognition, and object detection.¹⁴ Two networks, the encoder-decoder networks¹⁵ and generative adversarial networks (GANs)¹⁶ accelerated the use of deep learning in the image-to-image translation domain¹⁷ and image restoration.¹⁸ Specifically, in the nuclear medicine imaging field, the convolutional network has been used for image denoising, attenuation correction, and image reconstruction. Recently, deep learning methods have been demonstrated to be effective in cross-modality image-to-image translation, including the synthesis of CT images from MRI using deep convolutional neural networks,¹⁹ MR to CT translation with deep convolution adversarial networks,²⁰ β-amyloid PET to MRI image translation with GANs,²¹ MRI images to higher SNR PET images with the U-Net model,²² and generation of attenuation maps from SPECT emission data using deep learning models.²³

Inspired by cross-modality image-to-image translation, we proposed cross-tracer PET image translation methods using deep learning. Specifically, we observed that the level of glucose metabolism in brain regions is positively correlated with the amount of the SV2A,^24,25 which provides the possibility of predicting the PET image of ¹¹C-UCB-J from the PET image with ¹⁸F-FDG. Conversely, if some of the ¹¹C-UCB-J signals cannot be reproduced from ¹⁸F-FDG by this methodology, that may provide information about aspects of the ¹¹C-UCB-J and ¹⁸F-FDG signals that are independent.

In this work, we investigated a fully 3D U-Net architecture to predict the brain PET image of ¹¹C-UCB-J based on the brain PET image of ¹⁸F-FDG. In addition to investigating deep learning networks for static PET studies, we also trained the network for dynamic parametric PET images. We comprehensively optimized parameters and systematically investigated the quantification accuracy of ROIs in the brain using U-Net-based image prediction approaches. In addition, we explored a similar network training strategy to generate synthetic ¹¹C-PiB static images from ¹⁸F-FDG images.

2 |. MATERIALS AND METHODS

2.1 |. Human study dataset and brain imaging

As shown in Table 1, we included 54 subjects (25 males and 29 females; 33 Alzheimer’s disease [AD] and 21 cognitively normal [CN]; age 50–84). Similar datasets, as well as screening procedures and eligibility criteria, have been previously described.¹² Subjects underwent three brain dynamic PET scans of ¹⁸F-FDG, ¹¹C-UCB-J, and ¹¹C-PiB acquired on an HRRT scanner. The mean injected activity was ~5 mCi for ¹⁸F-FDG, ~16 mCi for ¹¹C-UCB-J, and ~13 mCi for ¹¹C-PiB. All the data were reconstructed with a motion-compensation list-mode reconstruction method,²⁶ in which OSEM was used with 2 iterations and 30 subsets. The original reconstructed images included 27 frames in 90 min after injection. The size of each image was 256 × 256 × 207 with a voxel size of 1.219 × 1.219 × 1.231 mm³.

TABLE 1.

Demographics and clinical diagnosis of dataset

Characteristic	Dataset (n = 54)
Age, mean (SD) [range], years	71.12 (8.07) [50.29–84.49]
Sex (Male: Female)	25:29
Weight, mean (SD) [range], kg	76.13 (16.24) [45–108.3]
18F-FDG Injection Activity, mean (SD) [range], mCi	4.85 (0.27) [3.74–5.15]
11C-UCB-J Injection Activity, mean (SD) [range], mCi	16.24 (4.71) [4.08–20.59]
11C-PiB-J Injection Activity, mean (SD) [range], mCi	12.58 (2.64) [4.78–14.82]
Diagnosis(CN:AD)	21:33

Abbreviations: AD, Alzheimer’s Disease; CN, cognitively normal.

For the static PET dataset, standard uptake value (SUV) images were generated from 60 to 90 min for ¹⁸F-FDG, 40 to 60 min for ¹¹C-UCB-J, and 50 to 70 min for ¹¹C-PiB. Then the SUVR image with the whole cerebellum as the reference was computed for all the SUV images. For the parametric PET dataset, ¹⁸F-FDG net uptake parameter K_i images were estimated with Patlak analysis using a population-based input function (PBIF). The time window for Patlak analysis was 30 to 90 min post-injection. K_i ratio images with the cerebellum as the reference region were calculated as (K_i/K_i[cerebellum]). Images of ¹¹C-UCB-J non-displaceable binding potential (BP_ND) were estimated with the Simplified Reference Tissue Model 2 (SRTM2) method by 90 min of data with centrum semiovale as the reference tissue.²⁷ Then, the distribution volume ratio with the cerebellum as the reference region (DVR_Cb) was calculated as (BP_ND + 1)/(BP_ND[cerebellum] + 1).

For each subject, a T1-weighted magnetic resonance imaging (MRI) scan was obtained at the Yale MRI Center for the registration between PET and MR images as well as the definition for regions of interest (ROI). Registration to MRI was implemented by the BioImage Suite software for the PET images.²⁸ Specifically, both the ¹⁸F-FDG images of SUVR and K_i Ratio and the ¹¹C-UCB-J images of SUVR and DVR of the same subject were registered with a rigid transformation to each individual’s MR image. The image size after registration was resliced to 176 × 240 × 256 with a voxel size of 1.2 × 1.055 × 1.055 mm³ in MR space for both ¹⁸F-FDG and ¹¹C-UCB-J images. Then, a brain mask derived from the aligned MR image was used to crop the images. ROIs were derived from the MR images using FreeSurfer version 6 (Massachusetts General Hospital).²⁹ Similarly, we registered ¹⁸F-FDG datasets (SUVR and K_i Ratio) and ¹¹C-PiB datasets (SUVR and K_i Ratio) to the MR space and generated ROIs from the MR image of each subject.

2.2 |. Deep learning network architecture

U-Net is a convolutional network architecture that has been widely applied to biomedical image processing. Unlike the fully convolutional network,³⁰ a large number of filters are also applied in the upsampling step in order to propagate context information to higher resolution layers.¹⁵

Since training such a 3D network based on entire images is computationally intensive, we extracted patches randomly from the images within the brain mask as the input and output. Moreover, the training data in terms of the number of patches are much larger than the number of images. The structure of the patch-based U-Net is shown in Figure 1. It consists of a contracting path (4 layers) and an expansive path (4 layers). U-net has concatenating skip connections between the contracting path and the corresponding expansive path, which makes up a U-shape. The contracting path contains a layer of a 3 × 3 × 3 convolution kernel followed by a rectified linear unit (ReLU) and a layer of 2 × 2 × 2 max pooling operation. The expansive path consists of a 2 × 2 × 2 up-convolution kernel and a 3 × 3 × 3 convolution kernel, followed by a ReLU. Symmetric padding was applied to maintain the same patch size after convolution. A fully convolutional layer was used to map the feature map to the label at the final stage. Thirty-two features were extracted in the first layer and were doubled at each downsampling step, then halved at each upsampling step as shown in Figure 1.

FIGURE 1.

The Architecture of The U-Net 3D model

2.3 |. Experimental setup

For ¹¹C-UCB-J image generation, two groups of studies were conducted to investigate the prediction accuracy of both SUVR and DVR images of ¹¹C-UCB-J using four networks, as summarized in Table 2. Two networks aimed to predict static ¹¹C-UCB-J SUVR images, where ¹⁸F-FDG SUVR images were used as the input data for Network #1 and ¹⁸F-FDG K_i Ratio images were used as the input data for Network #2. Another two networks were trained to predict parametric ¹¹C-UCB-J DVR images, where ¹⁸F-FDG SUVR images were used as the input data for Network #3 and ¹⁸F-FDG K_i Ratio images were used as the input data for Network #4. In ¹¹C-PiB PET image generation, ¹⁸F-FDG SUVR images were used as input, and ¹¹C-PiB SUVR images were used as output for Network #5.

TABLE 2.

Five networks used for ¹¹C-UCB-J and ¹¹C-PiB image generation

Network ID	Input image	Output image
	18F-FDG	11C-UCB-J	11C-PiB
#1	SUVR	SUVR	/
#2	K i Ratio	SUVR	/
#3	SUVR	DVR	/
#4	K i Ratio	DVR	/
#5	SUVR	/	SUVR

During the training, 240 epochs each containing 16000 patches were chosen for the convergence of the cost function. Each epoch consisted of 500 small batches. The patch size of 32 × 32 × 32 was applied in the ¹¹C-UCB-J image generation, while 64 × 64 × 64 was applied in the ¹¹C-PiB SUVR image generation. We chose the patch size based on a pilot study^31,32 (methods in Supplemental Materials), where we found the performances of 32 × 32 × 32 and 64 × 64 × 64 patches were similar for ¹¹C-UCB-J, while 64 × 64 × 64 patches outperformed 32 × 32 × 32 patches for ¹¹C-PiB image generation(Figure S1). The Adam optimizer was applied with an initial learning rate of 0.001 and the learning decay rate of 0.96 by each epoch. L₁ norm was used to calculate the loss between the generated patch and the target patch for the update since it showed better performance and faster convergence speed than L₂ during the training in this case. For testing, the patch size was set to 240 × 240 × 160 which is larger than the size of the input image to cover the whole image of each subject without overlap. Since we had 54 subjects, a “leave-11-out” cross-validation method was used, which means 43 subjects were used for training and the other 11 subjects were used for testing. Then we repeated this process four times until all the subjects had been used for testing. Subjects for both training and testing were selected randomly from AD and CN groups separately to form a balanced group with approximately the same ratio between the number of AD and CN subjects. All the experiments were carried out on the server at Yale PET Center with two GPUs (NVIDIA Quadra RTX 8000).

The proposed deep learning-based image synthesis approach was also compared to a simple fitting approach, based on the potential correlation between glucose metabolism and synaptic density. ¹¹C-UCB-J image prediction from ¹⁸F-FDG images using the polynomial fitting with the same dataset was applied. A population-averaged fitting curve was established for all brain voxels in ¹⁸F-FDG images as the input variables and the matching voxels in ¹¹C-UCB-J images as the target variables based on the same cross-validation strategy, which means the fitting curve was established based on 43 subjects, then the prediction was applied to 11 other testing subjects. This process was repeated four times until all the subjects were used for testing.

2.4 |. Statistical analysis

The normalized root mean square error (NRMSE) inside the brain between the predicted image and the true image was calculated as follows:

$$NRMSE=\sqrt{{\Vert I^{Pred}-I^{True}\Vert }^2/{\Vert I^{True}\Vert }^2},$$

where I^Pred is the predicted image from the ¹⁸F-FDG image and I^True is the true image of ¹¹C-UCB-J or ¹¹C-PiB. Structural similarity (SSIM) between the predicted image and the true image was calculated as follows:

$$SSIM=\frac{\left(2{\mu }_x{\mu }_y+c_1\right)\left(2{\sigma }_{xy}+c_2\right)}{\left({\mu }_x^2+{\mu }_y^2+c_1\right)\left({\sigma }_x^2+{\sigma }_y^2+c_2\right)},$$

where μ_x and μ_y are the mean values across all the voxels within the brain mask of the predicted and true images, σ_x and σ_y are the standard deviations of the predicted and true images across all voxels, and σ_xy is the covariance of the two images. c₁ and c₂ are two variables to stabilize the division with a weak denominator. By default, c₁ = (0.01 × L)², c₂ = (0.03 × L)², where L is the specified dynamic range value. Pearson’s correlation coefficient between the predicted image and the true image was calculated as follows:

$${\rho }_{X,Y}=\frac{cov\left(X,Y\right)}{{\sigma }_X{\sigma }_Y},$$

where X and Y are the vector representations of the predicted image and true image. For ROI-based quantification, the percent bias between the predicted image and the ground truth in a number of typical ROIs were calculated as follows:

$$Bias=\frac{{\mu }_{x,R}-{\mu }_{y,R}}{{\mu }_{y,R}}\times 100\%,$$

where R is the specific ROI as shown in Table 3.

TABLE 3.

Brain regions-of-interest

Index	ROI	Index	ROI
1	Hippocampus	10	Insula
2	Entorhinal	11	Thalamus
3	Precuneus	12	Caudate
4	Posterior Cingulate	13	Putamen
5	Parietal	14	Pallidum
6	Pons	15	Amygdala
7	Frontal	16	Cerebral white matter
8	Occipital	17	Anterior Cingulate
9	Temporal

Independent samples t-tests were applied between the CN and AD groups to test if there is a group difference in ROI percentage bias. A one-sample t-test was applied to check if the mean of ROI percentage bias is significantly different from zero for each network to compare the difference between predicted images and the true images. In addition, scatter plots with linear fitting were drawn to evaluate the consistency between the true images and the predicted images.

3 |. RESULTS

As shown in Figure 2, both using ¹⁸F-FDG SUVR and K_i Ratio images as the input can provide a satisfactory prediction of ¹¹C-UCB-J SUVR images. Visually, the predicted images appear to have less noise than the true images, and the predicted images are smoother than the true images. The smooth effect was also observed in other image denoising studies using the U-Net model.³² The U-Net model minimized the squared error between the predicted image and the true image, which may cause a smooth effect during prediction.³³ The mean NRMSE, SSIM, and Pearson coefficient across all AD and CN subjects for the ¹⁸F-FDG SUVR → ¹¹C-UCB-J SUVR network and ¹⁸F-FDG K_i Ratio → ¹¹C-UCB-J SUVR network are listed in Table 4. The SSIM and Pearson coefficient between the predicted ¹¹C-UCB-J SUVR image and the true image were statistically larger than that between the ¹⁸F-FDG input image and true ¹¹C-UCB-J SUVR for Network #1 and Network #2 according to paired-sample t-tests. The NRMSE of ¹⁸F-FDG SUVR → ¹¹C-UCB-J SUVR network is 16.2% ± 2.9% for the AD group and 15.6% ± 2.6% for the CN group, which outperformed the prediction results of ¹⁸F-FDG K_i Ratio → ¹¹C-UCB-J SUVR network. The SSIM and Pearson coefficient of ¹⁸F-FDG SUVR → ¹¹C-UCB-J SUVR network outperformed evaluation results of ¹⁸F-FDG K_i Ratio → ¹¹C-UCB-J SUVR network as well. Besides, using both ¹⁸F-FDG SUVR images and ¹⁸F-FDG K_i Ratio images as input could provide a reasonable prediction of ¹¹C-UCB-J DVR images. Using ¹⁸F-FDG SUVR image as the input provided superior NRMSE, SSIM, and Pearson coefficient across all the subjects as compared to using ¹⁸F-FDG K_i Ratio images as input as shown in Table 4.

FIGURE 2.

Representative predicted results of four networks including #1: ¹⁸F-FDG SUVR → ¹¹C-UCB-J SUVR, #2: ¹⁸F-FDG K_i Ratio → ¹¹C-UCB-J SUVR, #3: ¹⁸F-FDG SUVR → ¹¹C-UCB-J DVR, and #4: ¹⁸F-FDG K_i Ratio → ¹¹C-UCB-J DVR for (a) one Alzheimer’s disease subject and (b) one cognitively normal subject

TABLE 4.

NRMSE, SSIM, and Pearson correlation coefficients of predicted results compared to true ¹¹C-UCB-J images from four networks

	#1: SUVR to SUVR		#2: Ki Ratio to SUVR		#3: SUVR to DVR		#4: Ki Ratio to DVR
	Pred-True	Input-True	Pred-True	Input-True	Pred-True	Input-True	Pred-True	Input-True
AD
NRMSE	16.2% (2.9%)	/	17.8% (3.3%)	/	18.1% (3.3%)	/	20.0% (3.5%)	/
SSIM	0.903 (0.033)	0.866 (0.037)	0.880 (0.032)	0.737 (0.037)	0.890 (0.029)	0.773 (0.033)	0.875 (0.026)	0.795 (0.040)
ρ	0.963 (0.012)	0.932 (0.014)	0.955 (0.014)	0.884 (0.042)	0.937 (0.020)	0.863 (0.024)	0.924 (0.023)	0.799 (0.057)
CN
NRMSE	15.6% (2.6%)	/	16.8% (2.7%)	/	17.4% (2.4%)	/	19.0% (2.5%)	/
SSIM	0.910 (0.030)	0.874 (0.037)	0.890 (0.029)	0.751 (0.021)	0.902 (0.019)	0.781 (0.028)	0.885 (0.024)	0.807 (0.024)
ρ	0.967 (0.011)	0.944 (0.012)	0.961 (0.012)	0.908 (0.022)	0.945 (0.014)	0.883 (0.016)	0.935 (0.016)	0.835 (0.029)
All
NRMSE	16.0% (2.8%)	/	17.4% (3.1%)	/	17.9% (3.0%)	/	19.7% (3.2%)	/
SSIM	0.906 (0.032)	0.869 (0.037)	0.884 (0.031)	0.743 (0.033)	0.895 (0.026)	0.776 (0.032)	0.879 (0.026)	0.799 (0.035)
ρ	0.965 (0.011)	0.937 (0.014)	0.957 (0.013)	0.894 (0.037)	0.940 (0.018)	0.870 (0.023)	0.928 (0.021)	0.813 (0.051)

All the results were displayed by mean (SD)

Abbreviations: AD, Alzheimer’s Disease; All, All the Subjects; CN, Cognitively Normal.

Consistency analysis of histogram distribution was performed between the true ¹¹C-UCB-J image in relation to both the predicted ¹¹C-UCB-J image and the input ¹⁸F-FDG image across all four networks. As shown in Figure 3, for all the networks, the distribution of values in the predicted image is more consistent with that of the true image in terms of correlation as compared to the distribution of values in the input ¹⁸F-FDG image in both AD and CN subjects.

FIGURE 3.

Log-scaled joint histograms between predicted images and the ground truth, as well as between the input images and the ground truth for all the networks. The color represents the log-base 10 of the number of voxels with a given set of SUVR, DVR, or K_i Ratio values. AD refers to one Alzheimer’s disease subject and CN refers to one cognitively normal subject (same subjects as in Figure 2)

The predicted images based on the polynomial fitting method are shown in Figure 4. We observed that the texture of the predicted images is similar to the input ¹⁸F-FDG images in all the four image synthesis pairs, indicating that polynomial fitting could not help obtain additional structural features of ¹¹C-UCB-J images. We compared the overall quantification results of both deep learning network and polynomial fitting method as shown in Figure 5. For all the four image synthesis pairs, deep learning-based methods clearly showed smaller NRMSE and higher SSIM and Pearson coefficient indicating that the network could learn more structural information and outperforms the simple curve fitting method.

FIGURE 4.

Representative predicted results using polynomial fitting of four synthesis pairs including #1:¹⁸F-FDG SUVR → ¹¹C-UCB-J SUVR, #2: ¹⁸F-FDG K_i Ratio → ¹¹C-UCB-J SUVR, #3: ¹⁸F-FDG SUVR → ¹¹C-UCB-J DVR, and #4: ¹⁸F-FDG K_i Ratio → ¹¹C-UCB-J DVR for (a) one Alzheimer’s disease subject and (b) one cognitively normal subject

FIGURE 5.

Comparison between deep learning method (DL) and polynomial fitting (PF) curve method based on NRMSE(a, b), SSIM(c, d), and ρ(e, f) in both AD and CN groups. AD: Alzheimer’s Disease, CN: cognitively normal

The mean percentage bias and standard deviation of all the selected ROIs are shown in Table 5 and Table 6. As shown in Table 5, the mean percentage bias across all ROIs is −0.33% ± 7.43% in the AD group and −0.47% ± 7.28% in the CN group for Network #1 (¹⁸F-FDG SUVR → ¹¹C-UCB-J SUVR). For Network #2 (¹⁸F-FDG K_i Ratio → ¹¹C-UCB-J SUVR), the mean percentage bias is −0.68% ± 8.10% in the AD group, and −1.31% ± 6.99% in the CN group. In addition, the ROI bias of gray matter was calculated in both the AD and CN groups for global error estimation. The mean percentage bias is −1.61% ± 4.99% (p = 0.07) in the AD group and −1.33% ± 5.61% (p = 0.29) in the CN group for Network #1, while for Network #2, the mean percentage bias is −2.76% ± 4.96% (p < 0.05) in the AD group and −3.05% ± 5.33% (p < 0.05) in the CN group. The observed high standard deviation of ROI bias is secondary to tracer uptake variability across subjects.

TABLE 5.

ROI percent bias across AD and CN Groups for ¹⁸F-FDG SUVR to ¹¹C-UCB-J SUVR Network and ¹⁸F-FDG K_i Ratio to ¹¹C-UCB-J SUVR Network

ROI	18F-FDG SUVR → 11C-UCB-J SUVR		18F-FDG KiRatio → 11C-UCB-J SUVR
	AD	CN	AD	CN
Hippocampus	−0.29% (8.01%)	−1.50% (7.71%)	0.25% (10.3%)	−2.11% (7.17%)
Entorhinal	−2.84% (9.61%)	−0.89% (7.24%)	−2.90% (8.45%)	−3.02% (7.70%)
Precuneus	−2.17% (7.26%)	−1.38% (7.14%)	−4.32% (7.36%)*	−3.64% (6.65%)*
Cingulate Post	−1.50% (6.98%)	−1.15% (8.49%)	−3.05% (7.27%)*	−2.95% (6.86%)
Parietal	−2.13% (6.54%)	−1.17% (6.71%)	−3.78% (6.60%)*	−3.23% (6.37%)*
Pons	0.49% (7.34%)	−3.56% (6.15%)*	2.31% (9.17%)	−0.86% (7.71%)
Frontal	−1.60% (5.81%)	−2.00% (7.09%)	−3.15% (5.75%)*	−4.39% (6.97%)*
Occipital	−2.49% (6.75%)*	−1.88% (5.94%)	−3.15% (6.95%)*	−3.51% (5.64%)*
Temporal	−1.75% (6.73%)	−0.08% (6.42%)	−2.94% (6.40%)*	−2.03% (5.91%)
Insula	0.05% (6.63%)	−0.53% (6.66%)	−0.94% (7.34%)	−2.67% (5.93%)
Thalamus	−0.17% (6.73%)	−2.18% (6.32%)	1.06% (6.33%)	−1.77% (5.00%)
Caudate	1.10% (7.12%)	1.38% (6.77%)	−0.14% (7.71%)	0.20% (5.15%)
Putamen	2.33% (8.19%)	2.09% (7.71%)	2.22% (7.47%)	2.61% (6.93%)
Pallidum	3.04% (6.23%)*	−0.16% (6.63%)	3.83% (7.57%)*	2.20% (6.96%)
Amygdala	1.29% (9.59%)	2.96% (9.94%)	3.32% (11.0%)	3.50% (8.11%)
Cerebral White Matter	1.13% (6.32%)	1.37% (6.24%)	1.61% (7.15%)	0.87% (6.34%)
Cingulate Ant	−0.08% (7.60%)	0.66% (8.83%)	−1.79% (7.66%)	−1.51% (8.25%)
Gray Matter	−1.61% (4.99%)	−1.33% (5.61%)	−2.76% (4.96%)*	−3.05% (5.33%)*

All values are expressed as mean% ± SD%.

p values are obtained by a two-tailed one-sample t-test compared with the mean of zero.

^*p < 0.05.

TABLE 6.

ROI percent bias across AD and CN Groups for ¹⁸F-FDG SUVR to ¹¹C-UCB-J DVR Network and ¹⁸F-FDG K_i Ratio to ¹¹C-UCB-J DVR Network

ROI	18F-FDG SUVR → 11C-UCB-J DVR		18F-FDG Ki Ratio → 11C-UCB-J DVR
	AD	CN	AD	CN
Hippocampus	−0.37% (7.89%)	−3.55% (6.91%)*	−0.85% (9.39%)	−2.73% (6.95%)
Entorhinal	−1.95% (10.1%)	−1.43% (7.58%)	−0.99% (9.92%)	−2.70% (8.73%)
Precuneus	−4.01% (7.64%)*	−3.12% (6.57%)*	−5.34% (8.62%)*	−4.30% (7.56%)*
Cingulate Post	−3.17% (7.15%)*	−2.93% (7.61%)	−5.07% (8.60%)*	−3.94% (8.43%)*
Parietal	−2.87% (6.43%)*	−2.49% (6.75%)	−4.99% (7.44%)*	−3.82% (7.24%)*
Pons	−0.91% (9.07%)	−0.71% (7.71%)	10.12% (11.8%)*	5.88% (9.38%)*
Frontal	−1.90% (5.48%)	−3.61% (6.54%)*	−3.85% (6.34%)*	−4.59% (7.30%)*
Occipital	−2.81% (7.62%)*	−3.01% (6.27%)*	−4.14% (7.75%)*	−5.22% (6.91%)*
Temporal	−2.27% (6.69%)	−1.89% (5.87%)	−2.83% (6.50%)*	−2.84% (6.15%)*
Insula	−1.16% (6.66%)	−3.67% (6.14%)*	−1.15% (6.88%)	−4.37% (6.50%)*
Thalamus	−0.63% (6.12%)	−3.16% (7.10%)	−0.04% (6.56%)	−1.36% (7.22%)
Caudate	−0.43% (7.24%)	−1.39% (5.70%)	0.14% (7.73%)	−1.96% (5.50%)
Putamen	−0.30% (7.36%)	−1.05% (6.49%)	1.48% (7.77%)	−0.84% (6.39%)
Pallidum	0.33% (6.53%)	−2.03% (7.68%)	1.54% (7.23%)	−0.31% (7.21%)
Amygdala	0.03% (10.2%)	1.26% (8.62%)	3.64% (12.0%)	4.00% (8.69%)*
Cerebral White Matter	0.69% (6.26%)	0.91% (5.78%)	0.60% (7.76%)	1.15% (6.82%)
Cingulate Ant	−0.96% (6.61%)	−2.22% (6.96%)	−0.62% (7.06%)	−2.19% (7.38%)
Gray Matter	−2.31%(4.99%)*	−2.83%(5.27%)*	−3.56%(5.33%)*	−3.74%(5.76%)*

All values are expressed as mean% ± SD%.

p values are obtained by a two-tailed one-sample t-test compared with the mean of zero.

^*p < 0.05.

Two statistical comparisons were made. First, there were no statistically significant differences in bias between the AD and CN groups for any ROI. Within the hippocampus, for example, a region with significant ¹¹C-UCB-J group differences, p-values were 0.58, 0.32, 0.13, and 0.40 for Networks #1–4, respectively. Second, statistical analyses were performed to determine if mean biases across any group were significantly different from 0. As shown in Table 5, for Network #1, biases were significantly different from 0 within the occipital and pallidum ROIs in the AD group, and within the pons in the CN group, although the mean bias was <4% (in absolute value). The number of biased results is slightly larger in Network #2, indicating that the ¹¹C-UCB-J SUVR prediction network with ¹⁸F-FDG SUVR as the input leads to better evaluation performance. As shown in Table 6, the mean bias of all ROIs is −1.34% ± 7.47% in the AD group and −2.00% ± 6.88% in the CN group for Network #3 (¹⁸F-FDG SUVR → ¹¹C-UCB-J DVR), while −0.73% ±9.04% in the AD group and −1.70% ± 7.80% in the CN group for Network #4 (¹⁸F-FDG K_i Ratio → ¹¹C-UCB-J DVR). As with Networks #1 and #2, Networks #3 and #4 demonstrated no significant differences in ROI percentage bias between CN and AD groups. Regarding ROI mean biases significantly different from 0, Network #3 led to superior results in the pons, frontal, and temporal ROIs (one-sample t-tests, p > 0.05) in the AD group as compared to Network #4. Within the CN group, Network #3 network led to superior results in the posterior cingulate, parietal, pons, temporal, and amygdala ROIs (one-sample t-tests, p > 0.05) as compared to Network #4. The mean percentage bias of gray matter is −2.31% ± 4.99% (p < 0.05) in the AD group and −2.83% ± 5.27% (p < 0.05) in the CN group for Network #3, while for Network #4 is −3.56% ± 5.33% (p < 0.05) in the AD group and −3.74% ± 5.76% (p < 0.05) in the CN group.

A comparison between the deep learning method (DL) and polynomial curve fitting (PF) method on ROI bias in AD and CN groups is depicted in Figure 6. For polynomial results, the mean bias of all the ROIs is −3.78% ± 19.41% in the AD group and −1.34% ± 17.13% in the CN group for Network #1, while for Network #2 is −3.79% ± 20.17% in the AD group and −2.43% ± 18.05% in the CN group. Similarly, the mean bias of all ROIs is −5.29% ± 24.46% in the AD group and −2.47% ± 23.39% in the CN group for Network #3, while for Network #4 is −4.96% ± 25.29% in the AD group and −3.28% ± 24.55% in the CN group. For all four image synthesis pairs, the deep learning method shows smaller bias across most ROIs, indicating that the deep learning method is more applicable for cross-tracer image synthesis than the simple polynomial fitting method.

FIGURE 6.

Comparison between deep learning method (DL) and polynomial fitting curve (PF) method based on ROI percentage bias for ¹¹C-UCB-J SUVR prediction in the AD group (a) and the CN group (c) as well as ¹¹C-UCB-J DVR prediction in the AD group (b) and the CN group (d). AD, Alzheimer’s disease subjects; CN, cognitively normal subjects

Figure 7 depicted the mean correlation between the predicted image and true image across all subjects in three sample ROIs (hippocampus, precuneus, caudate) for all four networks. The results suggest that the predicted ¹¹C-UCB-J results are more consistent with those of the true ¹¹C-UCB-J results in the hippocampus and caudate. Across all networks, the r² values ranged from ~0.8 in the caudate, 0.5 to 0.6 in the hippocampus, and 0.4 to 0.6 in the precuneus, indicating that the prediction accuracy could be ROI dependent. This may reflect the biological differences between regions for SV2A and glucose metabolism in AD.

FIGURE 7.

Correlation of the mean values of three representative ROIs (hippocampus, precuneus, and caudate) across all the subjects for all the networks between ¹⁸F-FDG and ¹¹C-UCB-J. Blue triangles represent Alzheimer’s disease subjects, red dots represent cognitively normal subjects. The black line is the linear fitting of mean values across all the subjects of one specific ROI. The green line is the line of identity

4 |. DISCUSSION

In the ¹¹C-UCB-J image prediction study, we demonstrated the feasibility of generating synthetic ¹¹C-UCB-J PET images of synaptic density from ¹⁸F-FDG images. Since the two tracers have intrinsically different distributions and quantitation, we focused on generating the SUVR and DVR images, instead of absolute SUV images. Our results showed that using both ¹⁸F-FDG SUVR and K_i Ratio images can lead to a robust prediction of synthetic SUVR and DVR images of ¹¹C-UCB-J. The quantification bias of various ROIs is smaller than 8% for the ¹¹C-UCB-J SUVR image prediction and smaller than 10% for the ¹¹C-UCB-J DVR image prediction. Regarding input images, both SUVR and K_i Ratio images provided a similar prediction of ¹¹C-UCB-J static and parametric images. Using SUVR images as the input is slightly better in ROI analysis than using K_i Ratio images as input, potentially due to the lower noise in SUVR images. Since generating SUVR images required a much shorter acquisition time, it is more practical to use SUVR images as input as compared to using K_i Ratio images.

While the primary analyses focused on the generation of synthetic ¹¹C-UCB-J images from ¹⁸F-FDG data, an exploratory analysis investigated the possibility that ¹¹C-PiB images could be generated from ¹⁸F-FDG images with proper training using paired data. The predicted images of amyloid deposition (¹¹C-PiB SUVR) from ¹⁸F-FDG SUVR for a single subject with AD and normal cognition are depicted in Figure S2. Visual inspection of ¹¹C-PiB SUVR predicted images (Figure S2, row 3) demonstrates little overlap with the true ¹¹C-PiB SUVR image (Figure S2, row 2), in both AD and CN groups due to the very different spatial pattern of tracer uptake in both CN and AD subjects. Therefore, we incorporated an additional binary channel with information about the subject as CN or AD (health vs. disease) into the network. With the addition of this new channel (Figure S2, row 4), the predicted images are more consistent with the true ¹¹C-PiB SUVR images after incorporating diagnosis information. The distribution of the predicted ¹¹C-PiB SUVR image is more consistent with that of the true ¹¹C-PiB SUVR image for both AD and CN subjects according to Figure S3. The mean ROI bias is 0.34% ± 15.72% for AD subjects and −3.58% ± 14.44% for CN subjects according to Table S1. As significant amyloid deposition (as assessed by ¹¹C-PiB SUV) has been observed in the frontal, parietal, temporal, and occipital of patients with Alzheimer’s disease,² we compared the SUVR of both the predicted images and the true images for these typical ROIs to evaluate the physiological characteristics prediction accuracy as shown in Table S2. Overall, the mean SUVR of predicted ¹¹C-PiB images is consistent with the mean SUVR of true ¹¹C-PiB images in selected ROIs.

We have demonstrated the feasibility to predict ¹¹C-UCB-J or ¹¹C-PiB images from ¹⁸F-FDG images using the U-Net model. Both static and parametric images of ¹¹C-UCB-J generation have been achieved with reasonable accuracy. We also implemented ¹¹C-PiB static image prediction with the additional patient diagnosis information since the image distribution is quite different between CN and AD subjects. Importantly, as such diagnostic information may not be readily available during the initial evaluation of new patients, ¹¹C-PiB image prediction from ¹⁸F-FDG may be more useful in monitoring disease progression and evaluating response to therapy for research use.

Although we demonstrated the possibility of using deep learning methods to generate ¹¹C-UCB-J or ¹¹C-PiB images from ¹⁸F-FDG images, the synthetic resultant images cannot replace true images. The ¹¹C-UCB-J is not FDA approved for clinical use yet. However, these generated images can still potentially be of use in an AD research setting as an additional piece of diagnostic information, especially for research institutions where ¹¹C-UCB-J are not immediately readily accessible. Related to amyloid tracers, we do have FDA-approved ¹⁸F amyloid tracers for clinical use that could be used in locations without on-site cyclotrons, so there is a demand for generating ¹¹C-PiB images from ¹⁸F-FDG PET. Similar to ¹¹C-UCB-J, we expect the synthesized ¹¹C-PiB images could be used as a piece of additional information in the AD research study. Besides, the use of a single ¹⁸F-FDG scan will reduce total imaging sessions, radiation exposure, and costs for participants who need multiple scans.

This study has some limitations. First, the different noise levels of ¹¹C-UCB-J and ¹¹C-PiB images due to the differences in tracer uptake among the subjects after injection could increase the evaluation bias. Second, the training results could be improved via the inclusion of a larger dataset. For future work, we will attempt to generate other tracer images such as tau using ¹⁸F-FDG images. In addition, images from two tracers instead of only one could be used as the input to generate a third tracer image.

5 |. CONCLUSIONS

It is feasible to use fully 3D U-Net architecture to predict both ¹¹C-UCB-J and ¹¹C-PiB PET images from ¹⁸F-FDG images. Both ¹⁸F-FDG SUVR images and K_i Ratio images could be used as the input for the prediction of¹¹C-UCB-J static and parametric images. Using the SUVR image as input is slightly better in terms of evaluation accuracy as well as its accessibility. ¹⁸F-FDG SUVR images along with diagnostic information could be used as the input for the prediction of¹¹C-PiB SUVR images.

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available from the corresponding author upon reasonable request.