Generative diffusion model enables quantification of calibration‐free arterial spin labeling perfusion magnetic resonance imaging data in an Alzheimer's disease cohort

Abstract

INTRODUCTION

M0 images were missing in Siemens ASL data in Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset, prohibiting cerebral blood flow (CBF) quantification.

METHODS

A conditional latent diffusion model was trained and evaluated on in‐house datasets, then applied to the Siemens data in ADNI‐3. Regional CBF differences by Alzheimer's disease (AD) stages, their accuracy for AD classification, and CBF trajectory slopes were compared between generated data (Siemens) and acquired data (General Electric).

RESULTS

The diffusion model generated M0 images with high fidelity (SSIM = 0.918 ± 0.023, PSNR = 31.361 ± 2.537) and minimal CBF bias (mean difference is 0.21 ± 1.58 mL/100 g/min). Both generated and acquired CBF showed similar spatial patterns and decreasing trends with AD progression in specific AD‐related regions. Generated CBF also improved accuracy in classifying AD stages compared to qualitative perfusion images.

CONCLUSION

This study shows the potential of diffusion models for imputing missing modalities in large‐scale studies exploring the use of ASL as a biomarker of AD.

Highlights

• Using latent diffusion model, we can generate M0 image from control image in arterial spin labeling (ASL) with high fidelity.

• The generated M0 can be used for cerebral blood flow (CBF) quantification in Alzheimer's Disease Neuroimaging Initiative dataset.

• The performance of classification between Alzheimer's disease (AD) patients and cognitive normal people is better when using generated CBF maps than using non‐quantitative perfusion images.

• ASL CBF decreases with AD progression in key AD‐related brain regions.

1. INTRODUCTION

Alzheimer's disease (AD) is a progressive and devastating neurological condition that affects millions of individuals. Neuroimage biomarkers for the early detection of AD and/or mild cognitive impairment (MCI) play an important role in guiding preventative intervention and/or anti‐amyloid treatment for potential progression of AD. ¹ , ² , ³ , ⁴ Arterial spin labeling (ASL) is a non‐invasive method to quantitatively measure cerebral blood flow (CBF). ⁵ It has been shown that ASL CBF is highly consistent with fluorodeoxyglucose positron emission tomography (FDG‐PET), ⁶ which is one of the current clinical standards for AD diagnosis. A recent systematic review ⁷ summarizing 81 original studies reported widespread CBF reductions, particularly in temporoparietal and posterior cingulate regions in AD individuals; MCI was also associated with decreased CBF in the posterior cingulate. Further, the patterns of ASL CBF reductions are correlated with disease severity and progression of AD. ⁸ This demonstrates the potential use of ASL CBF as an imaging biomarker for the early diagnosis and prognosis of AD. ⁹

Alzheimer's Disease Neuroimaging Initiative (ADNI) (https://adni.loni.usc.edu/) is one of the pioneering open‐access datasets for studying AD that includes ASL in the research protocol. ADNI‐3 study included 3D ASL data from three major MR vendors. Calculating CBF requires both perfusion‐weighted images and a proton density or M0 image. However, the implementation of ASL techniques was different across vendors. Specifically, General Electric (GE) uses pseudo‐continuous ASL (pCASL) with M0 for CBF quantification, while Siemens, which consists of > 50% of the dataset, acquired 3D pulsed ASL (PASL) without the M0 image, prohibiting quantitative CBF calculation. This absence of M0 images in ADNI‐3 dataset greatly impairs the potential of using CBF as a biomarker for AD.

Deep‐learning (DL) based generative models, such as diffusion models ¹⁰ have shown the ability to generate high‐quality images. ¹¹ , ¹² In this work, we developed a conditional latent diffusion model (LDM) ¹³ to generate the M0 images for Siemens ASL data in ADNI‐3 to quantify CBF. Then, we compared the patterns of CBF variation with AD progression from generated (Siemens) and acquired (GE) CBF data with regional analyses. Three machine learning (ML) classifiers were applied to evaluate the capability of CBF and perfusion‐weighted images to classify different AD stages. Our work demonstrates that the proposed diffusion model can impute missing modality, thereby improving the power and precision for studying ASL CBF as a biomarker in AD.

2. MATERIALS AND METHODS

2.1. Study design

The framework of this study is illustrated in Figure 1. LDM ¹³ is a recently introduced technique that performs a diffusion process in the latent space to improve computational efficiency while providing a flexible conditioning mechanism. The model has an encoder, a latent‐space diffusion module, and a decoder. Prior to the forward diffusion process, the input image (M0) is first encoded into the latent space to reduce image dimension while not losing the perceptual content. The forward and backward diffusion process is conducted in the latent space. The output of the reverse diffusion process is fed to the decoder to convert the latent‐space image to the original image space. This encoder‐decoder pair was pretrained as in ¹³ and the model weights were fixed in this study. The conditioning mechanism is shown on the right side of Figure 1A. Since there is a direct relationship between the control and M0 images (see Supporting Information eq. 3), control images were chosen as the condition for generating M0 images. The condition is encoded with the same encoder and its latent‐space representation is concatenated with the latent‐space image in the intermediate representations of M0 in each step of the reversed diffusion process to perform the conditional probability propagation.

FIGURE 1.

Illustration of the conditional LDM and the inference process. (A) Model structure of the latent diffusion model. The image (M0) in the original pixel space is first encoded into the latent space with the E. Diffusion process is conducted in the latent space to improve the computational efficiency. The condition (control image) is also encoded into the latent space with the same encoder and concatenate with the latent space feature of M0. The reversed denoising network ${\epsilon }_{\theta }$ takes the time step t, latent space image $z_t$ and the latent‐space image of the condition ${\tau }_{\theta }(y)$ as input and produces the previous latent space image in the previous step. The final M0 image is produced with a decoder to recover the image from the latent space. (B) The sampling process (generation). The input images are first scaled with the histogram matching to match the histogram of the training data. The trained LDM generates 20 samples for each scan and averages to get a robust sample. The scale factor saved in the previous histogram matching is used to scale the image back to the original scale. E, encoder; LDM, latent diffusion model

2.2. In‐house dataset for model training and testing

The details of the imaging parameters of ADNI‐3 Siemens pulsed arterial spin labeling (PASL) include repetition time (TR) = 4000 ms, echo time (TE) = 20.26 ms, segmented 3D gradient and spin echo (GRASE) readout with a matrix size of 128 × 128 × 32 and a resolution of 1.9 × 1.9 × 4.5 mm, ³ TI = 2000 ms and TI₁ = 800 ms. ¹⁴ Two background suppression (BS) pulses were applied after the PASL inversion pulse. To build an M0 generation model from the control images, we collected a dataset (Dataset 1, N = 55) of paired data of (1) PASL with the same protocol as the ADNI‐3 Siemens protocol, and (2) manually disabled BS pulses and extended inversion time (TI) to 5s, as shown in Figure S1A,B. To improve model generalizability across the whole ADNI dataset, another dataset (Dataset 2, N = 58) was collected for fine‐tuning from another ADNI site. The details for datasets 1 and 2 are shown in Table S1. All subjects provided written informed consent according to a protocol approved by the Institutional Review Board.

RESEARCH IN CONTEXT

1. Systematic review: Previous studies of cerebral blood flow (CBF) variations in Alzheimer's disease (AD) using Alzheimer's Disease Neuroimaging Initiative Phase 3 (ADNI‐3) dataset cannot use part of the data as M0 images were missing due to vendor implementation. Deep Learning and generative models have the potential to generate this missing data.

2. Interpretation: This study applied diffusion model to generate the M0 image using control images in ASL. The model was trained and evaluated on in‐house datasets and then applied to the ADNI Siemens dataset to impute the missing M0 for CBF calculation. Both generated and acquired CBF data showed similar differentiation patterns by AD stages and decreasing trends with AD progression in specific AD‐related regions. Generated CBF also improved accuracy in distinguishing AD compared to qualitative perfusion data.

3. Future directions: Future studies can further explore the potential of using these generated data for larger scale studies on validating ASL as an imaging biomarker of AD.

Model training was implemented with Python and Pytorch. Thirty‐nine, 6, and 10 subjects from dataset 1 were randomly selected to be the training, internal testing, and external testing data, respectively. Forty, 5, and 13 of the 58 subjects in dataset 2 were selected to be training, internal testing, and external testing, respectively, and both datasets were combined during the fine‐tuning process to avoid catastrophic forgetting. ¹⁵ Training was performed on a Lambda cluster with NVIDIA 3090 graphics processing unit (GPU). Adam optimizer was used with initial learning rate of 2e‐6 and trained for 1000 epochs. Loss function of training the denoising network were the same as described in ¹³ ,

$$L_{LDM}:={\mathbb{E}}_{\mathcal{E}\left(x\right),\epsilon \sim \mathcal{N}\left(0,1\right),t}\left[\epsilon -{\epsilon }_{\theta }{\left(z_t,t,{\tau }_{\theta }\left(y\right)\right)}_2^2\right]$$ (1)

where $\epsilon$ is the noise added into the latent variable $z_t$ in diffusion step $t$. The model that achieved the best performance on the combined internal testing set was selected as the final model.

2.3. Model testing on combined dataset

For generative models, the generated images were sampled from a distribution. This enables the uncertainty measurement of the generation. We optimized the number of samples to average by comparing the averaged image of different number of samples to the reference standard image. The standard deviation map of the samples was used to evaluate the spatial distribution of the uncertainty. The quantitative evaluation of the model performance was conducted on combined testing dataset from dataset 1 and dataset 2. We calculated the similarity (normalized mean squared error [NMSE], peak signal‐to‐noise ratio [PSNR], and structural similarity index [SSIM]) between the generated M0 and the reference. To evaluate the accuracy of CBF quantification using generated M0 images, we calculated the same similarity metrics on CBF maps, and CBF bias in the whole brain, gray matter (GM) and white matter (WM) regions of interest (ROIs). The ROIs were segmented with SPM12 (fil.ion.ucl.ac.uk/spm/software/spm12/). To further validate whether our method complies with the physics model, we fitted the T1 maps from the generated M0 images and the control images which were compared with those fitted with acquired M0 and control images.

2.4. Application of LDM to ADNI dataset

Data used in the preparation of this article were obtained from ADNI‐3 database (adni.loni.usc.edu). The demographic information of the ADNI‐3 dataset is summarized in Table S2. All subjects were characterized into four diagnostic groups, including cognitively normal (CN), subjective memory concerns (SMCs), MCI, and AD. The sampling process is illustrated in Figure 1B. Control images of the ASL sequence were averaged and used as the condition. As image intensities vary across different sites, data were normalized using a histogram matching method. ¹⁶ Specifically, the histogram of the training data was averaged and compared to the mean GM and WM values to determine the percentile on the histogram to be used as the landmarks for normalization. The process of normalization can be described by the following equation:

$$I_{norm}=\frac{T_1+T_2}{V_{p_1}+V_{p_2}}\times I_{orig}$$ (2)

where $T_1$ and $T_2$ are learned landmark values for GM and WM, and $V_{p_1}$ and $V_{p_2}$ are the values of percentiles $p_1$ and $p_2$ on the histogram of the image to be normalized. This scale factor was calculated and saved for each individual data. After the generation of the M0 image, this scale factor was used to rescale the generated M0 back to the original scale.

For each subject in the ADNI‐3 dataset, a T1‐weighted structural MRI was acquired along with ASL with a resolution of 1×1×1mm³. To perform regional analysis, the generated M0 images and perfusion images were first coregistered to the T1 image and then normalized to the Montreal Neurological Institute (MNI) template space using SPM12. CBF maps were calculated and averaged for each diagnostic group, respectively. Automated anatomical labeling (AAL) template was used to get regional CBF values.

2.5. Regional analyses

Statistical analyses were conducted to compare regional differences among the four groups, as well as the difference between GE data and generated Siemens data. A generalized linear model was used to include an interaction term between diagnostic groups and scanner types. The interaction term represented the difference in CBF variation across AD stages between scanners. A non‐statistically significant interaction test indicated no statistically significant differences in CBF variation across CN, SMC, MCI, and AD groups between generated (Siemens) and acquired (GE) data. Since the data did not follow normal distribution, we conducted Wilcoxon ranking score transformations, thus all statistical tests conducted by comparing the ranking score (median) instead of mean. SAS 9.4 was used for statistical analyses.

2.6. ML analysis for binary classification

Three ML algorithms were used to build classifiers: Random Forest (RF), Real AdaBoost, ¹⁷ and Elastic Net. RF and AdaBoost are considered as non‐parametric approaches while Elastic Net is considered as parametric approach in case of strong linear predictors. For all classifiers, 10‐fold cross‐validation was used to evaluate model performance. We re‐iterated the learning process 10 times and applied the classifier to each of the testing samples. Thus, each study sample served as an independent testing case once. Receiver operating characteristic (ROC) curve was constructed using the predicted probability from 10 testing datasets combined and the area under the curve (AUC) with 95% confidence interval was used to assess prediction accuracy. SAS Enterprise Miner 15.1: High‐Performance procedures were used for ML.

3. RESULTS

3.1. Model evaluation on combined in‐house dataset

We generated multiple M0 images followed by averaging and plotted the curves of performance metrics (NMSE, SSIM, and PSNR) as a function of the number of averaged images. As shown in Figure 2, the quality of the averaged M0 image improved with more averages and stabilized after averaging 20 samples. The variation of performance metrics across all external testing subjects (indicated by the shaded area) illustrates the stability of the conditional LDM. Figure 2D–F shows one representative test case for an averaged M0 image, the reference standard, and the difference map between the two images, respectively. Figure 2G shows the standard deviation map across 100 generated samples. Most brain regions have low standard deviation, supporting the high certainty of the trained conditional LDM. The regions with relatively higher standard deviations correspond to cerebrospinal fluid (CSF) or vasculature, as red arrows indicate, which tend to have high fluctuations in MR images.

FIGURE 2.

Uncertainty evaluation of the diffusion model. (A–C) NMSE, SSIM, and PSNR between generated and RS M0 image with different numbers of averaged samples. The shaded area shows standard deviation across external testing cases. (D,E) A representative case of the reference standard and the generated image. (F) The difference map between the reference standard and the generated M0 image. (G) Standard deviation map across 100 generated samples. Red arrows show CSF areas where the standard deviation is higher than in other brain tissues. CSF, cerebrospinal fluid; NMSE, normalized mean square error; PSNR, peak signal‐to‐noise ratio; RS, reference standard; SSIM, structural similarity index measure

Based on these results, we chose to average 20 samples for each case in our following analysis to balance between generation time and image quality. We systematically compared the generated M0 and CBF maps to the reference standard images in the following aspects. First, a representative case of generated M0 and CBF images is shown in Figure 3A, both of which have high similarities to the reference standard. Second, the quantitative similarity metrics of the generated images are summarized in Figure 3B, demonstrating the excellent performance of the conditional LDM. The mean GM and WM CBF values of generated CBF maps and the reference standards showed a high consistency (r = 0.97). Finally, Figure 3C,D shows scatter plot and the mean difference between the generated CBF values and the reference standards in the whole brain, GM, and WM across all external testing cases. The mean difference is 0.21 ± 1.58 mL/100 g/min for the whole brain, 0.13 ± 1.83 mL/100 g/min for GM, and ‐0.61 ± 1.16 mL/100 g/min for WM, which is less than 5% of the corresponding CBF values. This shows our model's feasibility and fidelity in preserving the qualitative and quantitative features of the M0 image.

FIGURE 3.

Qualitative and quantitative model performance evaluation. (A) A representative case of the reference standard and generated M0 and CBF maps. Both M0 and CBF map produced by the diffusion model show high fidelity and similarity to the reference standard image. (B) Quantitative similarity metrics of the generated image and the reference standard including NMSE, SSIM, and PSNR. (C,D) Bias in CBF quantification measurements. (C) Scatter plot of the averaged CBF values of the reference standard and generated CBF images in both GM and WM. (D) Averaged CBF difference in the whole brain, GM and WM. Mean difference is 1.07 ± 2.12 mL/100 g/min for whole brain, 1.29 ± 2.51 mL/100 g/min for GM and ‐0.04 ± 1.44 mL/100 g/min for WM. CBF, cerebral blood flow; GM, gray matter; NMSE, normalized mean squared error; PSNR, peak signal‐to‐noise ratio; SSIM, structural similarity index measure; WM, white matter

We further evaluated the fidelity of our generated M0 images based on their consistency with the MR physics model. The relationship between M0 and the control image is dependent on the timing of BS pulses, as shown in Figures S1 and S2. We calculated the T1 map from the generated M0 image and the acquired control image and compared that with the reference standard. Figure S2C,D shows a high similarity between the T1 map calculated from the generated M0 and from the acquired M0. The largest variation between the generated T1 map and the reference standard T1 map occurs in the area containing blood vessels and CSF. Figure S2E,F shows the averaged T1 values of the reference standard and fitted from the generated M0. T1 values fitted from the generated M0 are close to the reference standard, and T1 values of both GM and WM are consistent with that reported in literature, ¹⁸ confirming the model preserves the MR physics properties.

3.2. Evaluation of the diffusion model on the ADNI dataset

Since the diffusion model was trained on a combined dataset acquired with the same vendor, and the imaging parameters were identical to those of the ADNI dataset, the acquired images were expected to have similar image contrast. Therefore, the trained model was directly applied to generate the missing M0 images for the Siemens ASL dataset in ADNI‐3. Similar to the previous steps, 20 samples were generated for each scan and averaged to produce the final M0 image for CBF calculation. CBF maps were calculated with the acquired perfusion images and the generated M0 images according to Eq. (1) (in Supporting Materials). Figure 4A–D shows the averaged perfusion (direct subtraction of label and control images) and CBF maps of each group for both Siemens and GE data, respectively. The perfusion maps are in arbitrary units, while the CBF maps are standardized to the unit of mL/100 g/min. A greater degree of spatial inhomogeneity can be observed in the perfusion maps compared to CBF maps, especially in top and bottom slices, which is likely due to spatial variations of coil sensitivities and different head positions. The CBF maps show improved homogeneity by normalizing the perfusion signal with a calibration scan (M0).

FIGURE 4.

Averaged perfusion and CBF maps for different groups in the ADNI dataset including CN, SMC, MCI, and AD. The perfusion maps show larger inhomogeneity across different brain regions, while CBF maps are more homogeneous. Both Siemens and GE data show a decreasing trend from CN/SMC to MCI to AD, especially in the CBF maps. Siemens pulsed ASL data show more vascular signals while GE pCASL data is smoother across the whole brain. AD, Alzheimer's disease; ADNI, Alzheimer's Disease Neuroimaging Initiative; CBF, cerebral blood flow; CN, cognitive normal; GE, General Electric; MCI, mild cognitive impairment; pCASL, pseudo‐continuous arterial spin labeling; SMC, subjective memory concern

3.3. Results of regional analysis

We first compared the regional CBF values of baseline visits in ADNI‐3 between generated data (Siemens) and acquired data (GE) across the four diagnostic groups using a generalized linear model with interaction between the diagnosis group and scanner types. The interaction test showed in 86 out of 90 ROIs, there are no statistically significant differences in CBF variation across CN, SMC, MCI, and AD groups between generated (Siemens) and acquired (GE) data. The details of statistical analysis can be found in Table S3. Based on the latest systematic review, ⁷ AD‐related CBF changes are most typically observed in the temporoparietal and posterior cingulate cortex (PCC) regions. Figure 5 shows boxplots of CBF values among the four groups in four representative ROIs, including PCC, precuneus, angular gyrus, and inferior temporal gyrus for both left and right hemispheres. Boxplots of CBF values in four other AD‐affected regions (cuneus, inferior parietal gyrus, supramarginal gyrus, and hippocampus on both hemispheres) are shown in Figure S3. Visual inspection concurred with the interaction test and showed a high level of similarity between generated and acquired CBF data, with a trend of decreasing CBF from CN to MCI and AD, while there is little CBF difference between CN and SMC. Siemens data had greater variances than GE data, as shown by larger interquartile ranges in the box plot.

FIGURE 5.

Regional analysis of the trend in different groups of the subjects. Four AD‐related ROIs are shown, including PCC, precuneus, inferior parietal gyrus, and angular gyrus. Data from Siemens and GE are shown in blue and red, respectively. Both data show similar trend in these ROIs, with a decreasing from CN/SMC to MCI to AD. AD, Alzheimer's disease; CN, cognitively normal; GE, General Electric; MCI, mild cognitive impairment; PCC, posterior cingulate cortex; ROI, region of interest; SMC, subjective memory concern

3.4. Results of ML classification

The performances of ML classifiers are shown in Table 1 and Table S4. The AUC of the ROC curve of three ML methods for 3 pairwise group classifications, including AD versus CN, AD versus MCI, and MCI versus CN are displayed. For both Siemens and GE data, the performance is above acceptable for clinical use ¹⁹ to separate AD from CN, with the best AUC of 0.77 95% confidence interval (CI): (0.59, 0.91) for Siemens data and 0.9 95% CI: (0.83, 0.98) for GE data, while the performance for classifying MCI and CN is below the clinically acceptable level of 0.7 for both Siemens and GE. The performance using perfusion features is lower than that using quantitative CBF features in AD versus CN, supporting the importance of using quantitative CBF rather than qualitative perfusion values for classification.

TABLE 1.

Performance of machine learning classification between AD and CN

AUC (AD vs. CN)	Ada Boost (95% CI)	Random Forest (95% CI)	Elastic Net (95% CI)
Siemens perfusion	0.67 (0.45, 0.9)	0.68 (0.46, 0.89)	0.61 (0.39, 0.83)
GE perfusion	0.81 (0.72, 0.91)	0.82 (0.71, 0.93)	0.5 (0.5, 0.5)
Siemens CBF	0.67 (0.47, 0.88)	0.77 (0.59, 0.91)	0.75 (0.63, 0.91)
GE CBF	0.9 (0.83, 0.98)	0.8 (0.63, 0.96)	0.84 (0.74, 0.94)

Note: AUC of theROC curve of the classification from three machine learning methods, including Ada Boost, Elastic Net, and Random Forest are shown. For both data from GE and Siemens, the performance of classification is better when using CBF used as features compared to using perfusion as features. The results from Siemens data are slightly lower than GE data.

Abbreviations: AD, Alzheimer's disease; AUC, area under the curve; CBF, cerebral blood flow; CI, confidence interval; CN, cognitively normal; GE, General Electric; ROC, receiver operating characteristic.

4. DISCUSSION

In this work, we demonstrated that the generative diffusion model trained on our in‐house dataset can be successfully applied to an existing public dataset (ADNI‐3) to generate M0 images conditioned on the control image for PASL acquisition. There is a direct and MR physics‐based relationship between control and M0 images, which justifies the use of diffusion model to generate M0 from acquired control images. With conditional LDM, we can achieve high‐quality and high‐fidelity M0 images without systematic bias in CBF quantification while preserving the consistency with the MR physics model. We also showed that the generated M0 image could be used for CBF quantification and prediction of CBF variations with AD progression, which has better performance than non‐quantitative perfusion analyses. The patterns of the generated CBF maps were validated by comparing with ASL dataset acquired from another vendor, which were consistent with the literature. The consistent pattern of decreasing CBF with AD progression in key AD‐related brain regions (temporoparietal, precuneus, and PCC) supports the use of ASL as an imaging biomarker for the diagnosis and monitoring of AD.

Recent developments in DL and generative models enable image‐to‐image translation in many different ways. For example, this can be achieved by training a neural network for direct mapping or using conditional GANs to generate a new modality from the input. ²⁰ In comparison, using the diffusion model offers advantages including better interpretability, shown by the uncertainty maps, and better flexibility to add conditions. In this work, we chose to generate the M0 image first and calculate the CBF map with the known kinetic model instead of directly generating CBF with perfusion and control image as the condition. This is because the SNR of the perfusion image is generally much lower than the control and M0 images, which means larger variations in the perfusion images, making the model less stable. This is confirmed with our preliminary study, ²¹ where the model that generated the M0 image followed by CBF calculation yielded better performance than the model that directly generated the CBF map. Furthermore, the step‐by‐step nature of this workflow enhanced interpretability for generating CBF maps.

In our study, we observed similar patterns of CBF reduction associated with AD progression from generated CBF data, which was consistent with previous studies, ⁸ , ⁹ and acquired ASL data from another vendor. However, there were still differences in regional CBF values between acquired GE and generated Siemens data, with the latter demonstrating larger variations across subjects. This is likely due to the difference between pCASL (GE) and PASL (Siemens) data. In general, with the identical inversion time (TI) and post‐labeling delay (PLD), pCASL yields higher SNR compared to PASL, as more labeled blood reaches the imaging slice by the time of acquisition, ²² which explains interindividual variations in PASL data. In addition, a greater amount of arterial transit effects indicated by intravascular signals can be observed in PASL images around the Circle of Willis and its major branches. The difference between pCASL and PASL protocols may also lead to different quantitative CBF values, although the overall CBF trends with AD progression are comparable between Siemens and GE data.

Traditional ML approaches are shown to have comparable performance with more complex DL methods for classification. ²³ In our studies, ML achieved overall good performance to classify AD from CN. The ML results showed reduced but acceptable accuracy in classifying AD from CN using generated data compared to the acquired data. This could be due to the small sample size (N = 9) of AD patients. For the performance from larger sample categories, for example, MCI versus CN, the performance gap is closer. Since the comparison between generated and acquired images was not based on the same cohorts, the comparison could be confounded by demographic factors and comorbidities, as well as different ASL techniques used.

There are some limitations of this study. First, there is no reference standard to be used as the reference for the generated M0 images in the ADNI‐3 Siemens dataset. In this study, we evaluated the fidelity of the proposed method by several validation methods, such as validating the T1 maps with the physics model and literature, as well as by comparing the detected patterns of CBF reductions across different groups between generated and acquired CBF data. Although this cannot completely replace the necessity of the reference standard image, this approach is suitable for real‐world scenarios where the dataset has already been acquired. Future studies could utilize datasets with existing references to validate further the use of generative models in large multi‐site datasets. Second, pCASL data has higher SNR than PASL, which is also indicated in our results, where GE pCASL data showed better performance in classification than Siemens PASL data. Future studies should consider this difference and further standardize the imaging protocol of ASL. Third, the ADNI dataset is imbalanced, with more CN and MCI individuals than AD, which may affect the statistical significance of group comparisons and the performance of ML classifications. Fourth, the comparison between Siemens and GE was based on different patient cohorts, which could be confounded by demographic factors and comorbidities.

5. CONCLUSION

In conclusion, we developed a conditional LDM for the generation of missing M0 image based on the control image in ASL acquisition, allowing CBF quantification for the ADNI‐3 dataset. The observed trend of decreasing CBF with AD progression supports the use of ASL as an imaging biomarker for the diagnosis and monitoring of AD.

CONFLICT OF INTEREST STATEMENT

Danny J. J. Wang is a co‐founder and shareholder of Hura Imaging, Inc. Other authors have no disclosures.

CONSENT STATEMENT

All subjects provided written informed consent according to a protocol approved by the Institutional Review Board.

Supporting information

Supporting Information

Supporting Information

Shou Q, Cen S, Chen N‐K; for the Alzheimer's Disease Neuroimaging Initiative . et al., Generative diffusion model enables quantification of calibration‐free arterial spin labeling perfusion magnetic resonance imaging data in an Alzheimer's disease cohort. Alzheimer's Dement. 2025;17:e70214. 10.1002/dad2.70214

DATA AVAILABILITY STATEMENT

The data supporting the findings of this study are publicly available through the Alzheimer's Disease Neuroimaging Initiative (ADNI) at https://adni.loni.usc.edu. Code for this paper is available on GitHub: https://github.com/qinyangshou/M0_generation_ADNI.