Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Proc SPIE Int Soc Opt Eng . Author manuscript; available in PMC: 2026 Apr 11. Published in final edited form as: Proc SPIE Int Soc Opt Eng. 2025 Apr 11;13406:1340621. doi: 10.1117/12.3047135 Search in PMC Search in PubMed View in NLM Catalog Add to search A technical assessment of latent diffusion for Alzheimer’s Disease progression Elyssa McMaster Elyssa McMaster a Department of Electrical and Computer Engineering, Vanderbilt University, Nashville, TN, USA Find articles by Elyssa McMaster a, * , Lemuel Puglisi Lemuel Puglisi b Department of Computer Science, University of Catania, Catania, Italy Find articles by Lemuel Puglisi b , Chenyu Gao Chenyu Gao a Department of Electrical and Computer Engineering, Vanderbilt University, Nashville, TN, USA Find articles by Chenyu Gao a , Aravind R Krishnan Aravind R Krishnan a Department of Electrical and Computer Engineering, Vanderbilt University, Nashville, TN, USA Find articles by Aravind R Krishnan a , Adam M Saunders Adam M Saunders a Department of Electrical and Computer Engineering, Vanderbilt University, Nashville, TN, USA Find articles by Adam M Saunders a , Daniele Ravi Daniele Ravi c School of Physics, Engineering and CS, University of Hertfordshire, UK Find articles by Daniele Ravi c , Lori L Beason-Held Lori L Beason-Held d Laboratory of Behavioral Neuroscience, National Institute on Aging, National Institutes of Health, Baltimore, MD, USA Find articles by Lori L Beason-Held d , Susan M Resnick Susan M Resnick d Laboratory of Behavioral Neuroscience, National Institute on Aging, National Institutes of Health, Baltimore, MD, USA Find articles by Susan M Resnick d , Lianrui Zuo Lianrui Zuo a Department of Electrical and Computer Engineering, Vanderbilt University, Nashville, TN, USA Find articles by Lianrui Zuo a , Daniel Moyer Daniel Moyer e Department of Computer Science, Vanderbilt University, Nashville, TN, USA Find articles by Daniel Moyer e , Bennett A Landman Bennett A Landman a Department of Electrical and Computer Engineering, Vanderbilt University, Nashville, TN, USA e Department of Computer Science, Vanderbilt University, Nashville, TN, USA Find articles by Bennett A Landman a, e Author information Article notes Copyright and License information a Department of Electrical and Computer Engineering, Vanderbilt University, Nashville, TN, USA b Department of Computer Science, University of Catania, Catania, Italy c School of Physics, Engineering and CS, University of Hertfordshire, UK d Laboratory of Behavioral Neuroscience, National Institute on Aging, National Institutes of Health, Baltimore, MD, USA e Department of Computer Science, Vanderbilt University, Nashville, TN, USA * Email: [email protected] Issue date 2025 Feb. PMC Copyright notice PMCID: PMC12726967 NIHMSID: NIHMS2049059 PMID: 41445917 The publisher's version of this article is available at Proc SPIE Int Soc Opt Eng Abstract Latent Diffusion Models (LDMs) introduce exciting opportunities in medical imaging, from disease progression prediction to interpolation to generate entire datasets of rare data. The stochastic nature of generative models makes it challenging to validate their outputs and assess their robustness across diverse datasets. BrLP, a state-of-the-art LDM for T1-weighted (T1w) images that incorporates auxiliary brain volume information, has been evaluated on Alzheimer’s Disease (AD) progression, and achieves structural similarity index (SSIM) of 0.91 ± 0.03. In this work, we conducted a pilot study of the BrLP model using the Baltimore Longitudinal Study of Aging (BLSA) dataset. Our objectives are to (1) evaluate the model performance on an external dataset using pretrained image-based and brain image-based metrics such as mean squared error (MSE), similarity index, and mean absolute error (MAE) between conditional and unconditional brain regions; and (2) determine if a harmonization step, in addition to the proposed model’s preprocessing steps, is required to improve performance. We found that the BrLP is robust to T1w imaging scanner effects, and harmonization is not required. However, there exists a bias toward a younger population compared to the BLSA cohort (BrLP reported 0.91 ± 0.03 SSIM in a cohort with age range 74 ±7 years; SSIM in our experiment is 0.91 ± 0.03 in cognitively normal subjects with age range 79.47 years ± 7.35 and AD cohort 0.90 ± 0.012 with age range 83.73 years ± 6.01). Interestingly, when the model’s input was changed to simulate AD progression instead of normal aging, a higher SSIM of 0.91 ± 0.0012 was achieved compared to a ground truth non-AD scan, suggesting a potential mismatch. However, the resulting lower conditional volume regions function as expected. The model’s architecture shows promise for longitudinal T1w imaging studies. Keywords: Latent diffusion model, T1-weighted imaging, Alzheimer’s Disease 1. INTRODUCTION T1-weighted (T1w) imaging is a widely used magnetic resonance technique that provides qualitative structural information of the brain [ 1 ]. During the disease progression of Alzheimer’s Disease (AD), the brain undergoes substantial structural changes that can be observed with longitudinal T1w imaging data [ 2 ]. Even though the brain changes during AD progression, the acquisition of scans from subjects with AD and quantification of these structural changes still challenge researchers due to the heterogenous nature of disease progression and data collection. People with AD progress from cognitively normal presentation to mild cognitive impairment (MCI) to AD, each with individualized time scale, symptom presentation, and structural brain changes [ 3 ]. Large-scale longitudinal studies involve multiple scans of the same person at different time points (often separated by years). These studies have provided a wealth of information about the aging process and the progression of different disease [ 4 ], [ 5 ], [ 6 ]. In a large-scale longitudinal study, however, we can only see the brain at fixed points in time, and we must interpolate the changes in the brain’s structure between scans to learn about the changes that occur during the progression of disease ( Figure 1 ). AD presents a unique challenge as progression varies from person to person, population to population, and between types of AD, and these effects have not been well-characterized [ 7 ]. Figure 1: Open in a new tab Reliable synthetic images can provide a solution to incomplete disease progression datasets by interpolating MRI images between scans separated by multiple years. Here, the principles of image registration and interpolation combine with latent diffusion interpolation, but the integration of two forms of interpolation present a challenge for image processing. Understanding the progression of AD and its effect on brain structure has evolved into a major medical and technical problem. With the growth of an aging population, there exists demand to use state-of-the-art computer vision techniques to analyze the changes in individual anatomy in vivo during disease progression. This analysis contributes fundamental knowledge about the aging process and structural deviations caused by AD. The recent advancements in generative AI, and specifically in latent diffusion models (LDMs) [ 8 ], have the potential to address the lack of data that prevents the study of a particular biological phenomenon by generating reliable synthetic images. Reliable generative AI can help us to understand the mechanisms of AD both on a large-scale fundamental level and on an individual level [ 9 ]. Leveraging the limited longitudinal data available to model the progression of AD is a significant technological advancement. If sufficiently reliable, this approach could improve understanding of the disease and eventually, patient outcomes. Generative AI models have attempted to model the structural progression of AD to address this medical problem [ 9 ], [ 10 ], [ 11 ]. A recent development in generative AI, the LDM, encodes the latent space necessary to use a Denoising Diffusion Probabilistic Model (DDPM) [ 8 ], [ 12 ]. The diffusion model takes training data and incrementally adds Gaussian noise until the image consists of pure noise and then uses a neural network to predict the noise and generate a new image [ 12 ]. Such a model can limit memory necessary to generate synthetic images with the use of a latent space and meaningful representations of human brains. The first architecture to accommodate both issues of computer memory and human anatomy is Brain Latent Progression (BrLP). BrLP is a latent diffusion model that is the first of its kind to leverage prior knowledge of progression of disease with the insertion of an auxiliary model in the inference process [ 9 ]. Prior to BrLP, no individual pipeline combined conditioning on subject metadata and the use of longitudinal data with continuous spatiotemporal progression that is reasonable with accessible contemporary technology. BrLP leverages observed brain volume changes due to AD and applies them to their predictions of individual anatomy at different disease statuses. As a latent diffusion model, the BrLP pipeline first trains an autoencoder and trains a UNet as based on the latent information learned from the autoencoder. It uses a ControlNet [ 13 ] with the LDM to apply the latent information learned about different disease statuses to individual anatomy. The ControlNet is trained based on longitudinal data, allowing BrLP to learn from spatiotemporal input in large-scale studies. BrLP exploits Disease Course Mapping (DCM) [ 14 ], [ 15 ] to learn longitudinal volumetric trajectories of key AD-related brain regions, subsequently integrating this knowledge to inform the image generation process. The inference process uses Latent Average Stabilization (LAS), a technique developed with the BrLP model, to handle irregular patterns in aging and disease progression. In this work, we aim to assess the performance of BrLP on an external dataset, the Baltimore Longitudinal Study of Aging (BLSA) [ 20 ], [ 21 ]. By taking an external dataset, we evaluate if the model is susceptible to acquisition difference and evaluate if the model requires harmonization beyond the preprocessing procedure outlined in the BrLP paper [ 9 ]. This evaluation helps us determine if the model is ready to be applied out-of-the-box to maximize the applicability to rare accessible longitudinal AD data. Additionally, this paper provides a detailed analysis to assess the strengths and weakness of the current state-of-the-art. To assure maximum translational and clinical impact, we aim to articulate the model’s reproducibility for major Alzheimer’s studies with controls and AD subjects, and then test the model’s ability to condition based on disease status by applying an Alzheimer’s status to cognitively normal subjects to test its sensitivity to conditions. 2. METHODOLOGY The dataset collected for the BrLP study included 2,805 subjects across three large-scale, publicly available longitudinal studies: The Alzheimer’s Disease Neuroimaging Initiative (ADNI) [ 16 ], The Australian Imaging, Biomarkers, and Lifestyle (AIBL) [ 17 ] study of aging, and OASIS-3 [ 18 ]). Among the 2,805 subjects, 11,730 available T1w scans comprised the longitudinal dataset, with average age 74 ± 7 years. The data were preprocessed the same way regardless of dataset and acquisition scanner and randomly split into training, validation, and testing (80/5/15) without overlap. Subjects were grouped by cognitive status at the end of the study, where 43.8% of subjects maintained cognitively normal status, 25.7% show MCI, and 30.5% show AD. All images are resampled to 1.5 mm 3 isotropic resolution before training and inference. BrLP, with the auxiliary model for brain region volume and the advent of a new spatiotemporal approach, achieves 0.91 ± 0.03 SSIM and 0.004 ± 0.002 MSE in the cohorts it trained on [ 9 ]. The architecture as published was compared to the base model for latent diffusion and the base model combined with either the auxiliary model or the LAS technique, and the experiments show the best performance from the combination of all three and the evolution of spatiotemporal generative image modeling. BrLP has also outperformed DaniNet [ 11 ] and CounterSynth [ 10 ] in similarity metrics with ground truth images in a single-image/non-longitudinal study, where the auxiliary model is Linear Model, as well as Latent-SADM (a latent diffusion implementation of Sequence-Aware Diffusion Model for Longitudinal Medical Image Generation, SADM) [ 19 ] in a longitudinal study with DCM. BrLP has shown consistent and repeatable generation of individual anatomy conditioned on disease progression in AD datasets compared to baseline models, but it has not yet been validated on external longitudinal datasets or shown quantifiable changes on controls conditioned for disease in the model. We aim to evaluate the design of the original BrLP study on BLSA to test the current iteration of the model’s reproducibility in longitudinal AD studies, but in a cross-sectional context with input of only one image per prediction. A cross-sectional study allows for us to see how impactful this approach can become for extremely small datasets and without the benefit of longitudinal data to inform the diffusion model. In a proof-of-concept evaluation, we select 30 subjects (15 cognitive normal and 15 AD) with longitudinal T1w scans from BLSA. Though there are thousands of subjects included in the BLSA, few of them have (a) an AD diagnosis (as opposed to MCI or other unrelated cognitive impairment) by the end of their sequence of acquisitions; (b) images acquired from scanners similar enough to avoid a harmonization issue in our dataset, and (c) at least two scans that fit the scanner criteria. Our subjects have age range 69 years to 94 years, mean 81.6 years ± 6.94. All scans were acquired on a Philips 3T scanner with a magnetization-prepared rapid gradient echo (MPRAGE) sequence to minimize site effects between longitudinal scans. Acquisition resolution is 1 mm 3 isotropic voxels. The AD subjects were selected based on two available longitudinal data where AD is present by the last visit used in our study (though not necessarily at the first, where they may have been cognitively normal or present MCI), which is consistent to the labeling process of the original BrLP study [ 9 ]. Due to the higher availability of cognitively normal participants, the control cohort were selected based on the same criteria, but randomly within those criteria, resulting in AD cohort average age 83.73 years ± 6.01 and normal cohort age 79.47 years ± 7.35. We preprocessed the data with the same pipeline as the original BrLP study [ 9 ], [ 22 ] for all scans in our dataset, which includes N4 bias field correction from ANTs [ 23 ], skull stripping via SynthStrip from FreeSurfer [ 24 ], affine registration to MNI 152 space with SyN from ANTs [ 25 ], brain region segmentation with SynthSeg from FreeSurfer [ 26 ], and WhiteStripe intensity normalization [ 27 ]. We then ran the inference process with the model weights from BrLP (paper published May 2024, weights available at MICCAI 2024) for each dataset and generated synthetic images ( Figure 2 ). Figure 2: Open in a new tab In the cross-sectional/single-image context, a Linear Model (LM) is employed to inform the reverse diffusion process. Since this is a cross-sectional evaluation of a longitudinal study, we have a ground truth image to compare the model’s output to by three metrics: structural similarity index (SSIM), mean square error (MSE), and volume mean absolute error (MAE). The BrLP inference process accepts several hyperparameters with its input to assure meaningful representations of disease progression. Every subject’s metadata includes the original scan and segmentation, subject and scan session ID, age at the time of that scan, and sex. When running the model, it accepts the metadata information, as well as target age, target diagnosis, and the number of images to generate between the original image and target age. We generate an image for every year between input age and target age for consistency and to observe the predicted changes based on the conditioning of the model. After the synthetic images had been produced, we ran SynthSeg on the new synthetic images to compare brain region volumes between ground truth and synthetic longitudinal data with the same age as the following ground truth scan. We performed the same steps for the cognitively normal and AD cohorts. To qualitatively and quantitatively understand the model’s expectations for brain volume changes introduced by AD, we added another cohort, titled ‘Add AD’, that conditioned the cognitively normal group to produce synthetic images of their brains with Alzheimer’s in the same number of years as their ground truth normal synthetic image. We then compared the cognitively normal longitudinal images to the synthetic AD images to understand expected change in the same anatomy with the onset of Alzheimer’s. This way, we can compare the aging process in both synthetic cognitively normal and synthetic AD scans to a cognitively normal ground truth to understand the conditions driving the model when under different circumstances. The initial BrLP study uses image-based and volume-based metrics to quantify and compare the ground truth and synthetic images. The study uses SSIM and MSE to evaluate the quality of the synthetic images on an image-wide basis. In its architecture, BrLP conditions some brain regions based on a function of brain volume from pathology, and checks results based on unconditioned regions of the brain. The model conditions the hippocampus, amygdala, and lateral ventricle, and its original paper includes the thalamus and CSF as unconditional volumes for analysis, and use mean absolute error (MAE) as the metric to evaluate the model’s generation of those brain regions [ 9 ]. We evaluate the model’s performance on our dataset with the same metrics. The data are resampled from acquisition resolution to BrLP’s selected 1.5 mm 3 isotropic resolution for registration and evaluation. 3. RESULTS The quantitative results of our tests showed consistent results between the original BrLP study and our cognitively normal cohort from BLSA, as we achieved SSIM of 0.91 ± 0.01 and MSE of 0.0055 ± 0.0011, which is reasonably close to the reported SSIM of 0.91 ± 0.03 and MSE of 0.004 ± 0.002 (see Figure 3 ) [ 9 ]. The Alzheimer’s Disease cohort achieved a SSIM of 0.90 ± 0.012, and the normal cohort with AD-conditioned longitudinal scans achieved 0.91 ± 0.01 ( Figure 3 ). The image-based metrics show us consistent results between the original BrLP study and the external dataset of BLSA. Figure 3: Open in a new tab Our quantitative results compare each cohort’s synthetic images to their same-age ground truth counterparts. We observe consistent reproducibility of the BrLP paper’s results in image-based and volume-based metrics. The higher MAE in the lateral ventricle and CSF of the ‘Add AD’ cohort paired with the low MSE and high SSIM show the model’s ability to condition based on diagnosis in the brain volume regions while maintaining personalized anatomy. The variance in the segmentation of the conditional and unconditional regional volumes show an exciting advancement in synthetic brain imaging and disease progression modeling. The normal cohort’s conditional and unconditional regions have MAE like those in the original paper, and in some cases, show lower MAE -- possibly due to the smaller population size. The Alzheimer’s Disease cohort showed less consistently repeatable results, especially in the conditional lateral ventricle and unconditional CSF MAE metrics. Subjects in the normal cohort with synthetic AD condition showed higher image similarity than the ground truth AD cohort ( Figure 3 ), but they did not show the same consistency as the synthetic normal image to the ground truth normal image. The high image-based similarity of the normal cohort with synthetic AD images shows promise in the model’s ability to generate personalized anatomy with volumetric changes like those that a brain that changes status from cognitively normal to AD would undergo. In both the conditional and unconditional regions of interest, we see a bias toward the synthetic normal image when we compare mean absolute error of the brain region volumes, which indicates that this feature of the model adjusts volumetric expectations based on disease status. Our qualitative results show the visual differences between the ground truth and synthetic images ( Figure 4 ). The ground truth images at 1 mm 3 isotropic are resized to 1.5 mm 3 isotropic for BrLP inference. This resizing results in a smoother synthetic image compared to the original ground truth images. The synthetic images struggle to characterize the organic boundaries of regions of interest in the gray matter/white matter boundary visible in the ground truth images, even when present in the image input into the model. Although the model successfully generates images with personalize anatomy with high SSIM and MSE scores, the smoothing introduced during the generation process leads to a loss of some individual features, such as lesions, that may be critical for diagnosing AD cases ( Figure 4 ). Figure 4: Open in a new tab We see the high SSIM score reflected in the generation of personalized anatomy, but several cases where the model smooths individual lesions for lower similarity. 4. DISCUSSION Though the model shows substantial similarity between brain volumes of ground truth and synthetic images, the qualitative images show high smoothness in the gray matter/white matter boundary, as seen in Figure 4 . Studies have shown that AD has a large impact on white matter in addition to brain volume measurements [ 28 ], so advancements in the generation of images with accurate gray matter/white matter boundaries could further the study of AD. BrLP shows promise for consistent repeatable results, which will change the study of the progression of structural images in AD. The consistent results suggest that the model can handle an external dataset, given that the images are preprocessed with the same pipeline before running inference, without a complicated harmonization step to minimize site effects. Since the model shows subtle changes between cognitively normal and AD within the same anatomy, it shows a promising step in the architecture necessary to evaluate different diseases and disorders, possibly with different modalities and different cohorts. It is important to note the brain regions that auxiliary model uses to condition the model may be worth changing depending on cohort or disease, as Alzheimer’s Disease notably impacts the volume of the hippocampus, and the aging process widens the ventricles over time [ 29 ]. In conclusion in this study, we identified that BrLP is stable on an external dataset, and the preprocessing steps defined by the authors are robust enough to avoid a complicated harmonization step. We recognize that our study has several limitations. First, there were limitations on the number of AD subjects with longitudinal data that fit our criteria in BLSA. It is possible that a larger cohort (> 15 subjects per normal and AD cohorts) may have shown results more like the original BrLP paper. The BLSA cohort (especially within the AD cohort) had a significantly higher mean age (83.73 years ± 6.01) than the subjects in the initial BrLP study (74 years ± 7), which may have led to lower SSIM in the AD cohort from BLSA – so high that the mean age of the AD cohort represents data from outside of one standard deviation above the mean from the BrLP study. It is possible that the model is biased toward some kind of earlier onset Alzheimer’s Disease variant compared to the presentations within the participants in the BLSA study. BrLP’s architecture presents an opportunity to leverage T1w imaging data of Alzheimer’s Disease. The prior knowledge and inference designs of the model outperform baselines and continue to work on external datasets. The reproducibility of the model shows potential across imaging modalities and longitudinal studies. ACKNOWLEDGEMENTS The Vanderbilt Institute for Clinical and Translational Research (VICTR) is funded by the National Center for Advancing Translational Sciences (NCATS) Clinical Translational Science Award (CTSA) Program, Award Number 5UL1TR002243-03. This work was conducted in part using the resources of the Advanced Computing Center for Research and Education at Vanderbilt University, Nashville, TN. ADSP U24AG074855, NIH 1R01EB017230 Uncertainty in Diffusion MRI. The BLSA is supported by the Intramural Research Program, National Institute on Aging, NIH. This work was supported by National Cancer Institute (NCI), Grant/Award Number: R01 CA253923 and P50HD103537 VKC. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH. Lemuel Puglisi is enrolled in a PhD program at the University of Catania, fully funded by PNRR (DM 118/2023). We used generative artificial intelligence (AI) to create code segments based on task descriptions, as well as to debug, edit, and autocomplete code. Additionally, generative AI technologies have been employed to assist in structuring sentences and performing grammatical checks. The conceptualization, ideation, and all prompts provided to the AI originated entirely from the authors’ creative and intellectual efforts. We take accountability for the review of all content generated by AI in this work. REFERENCES [1]. Howarth C, Hutton C, and Deichmann R, “Improvement of the image quality of T1-weighted anatomical brain scans,” Neuroimage, vol. 29, no. 3, 2006, doi: 10.1016/j.neuroimage.2005.08.004. [ DOI ] [ Google Scholar ] [2]. Sattari N, Faeghi F, Shekarchi B, and Heidari MH, “Assessing the Changes of Cortical Thickness in Alzheimer Disease With MRI Using Freesurfer Software,” Basic Clin Neurosci, vol. 13, no. 2, 2022, doi: 10.32598/bcn.2021.1779.1. [ DOI ] [ Google Scholar ] [3]. Davatzikos C, Bhatt P, Shaw LM, Batmanghelich KN, and Trojanowski JQ, “Prediction of MCI to AD conversion, via MRI, CSF biomarkers, and pattern classification,” Neurobiol Aging, vol. 32, no. 12, 2011, doi: 10.1016/j.neurobiolaging.2010.05.023. [ DOI ] [ Google Scholar ] [4]. Resnick SM, Pham DL, Kraut MA, Zonderman AB, and Davatzikos C, “Longitudinal magnetic resonance imaging studies of older adults: A shrinking brain,” Journal of Neuroscience, vol. 23, no. 8, 2003, doi: 10.1523/jneurosci.23-08-03295.2003. [ DOI ] [ Google Scholar ] [5]. Chatterjee P et al. , “Plasma Aβ42/40 ratio, p-tau181, GFAP, and NfL across the Alzheimer’s disease continuum: A cross-sectional and longitudinal study in the AIBL cohort,” Alzheimer’s and Dementia, vol. 19, no. 4, 2023, doi: 10.1002/alz.12724. [ DOI ] [ Google Scholar ] [6]. Kruggel F, Masaki F, and Solodkin A, “Analysis of longitudinal diffusion-weighted images in healthy and pathological aging: An ADNI study,” J Neurosci Methods, vol. 278, 2017, doi: 10.1016/j.jneumeth.2016.12.020. [ DOI ] [ Google Scholar ] [7]. Deture MA and Dickson DW, “The neuropathological diagnosis of Alzheimer’s disease,” 2019. doi: 10.1186/s13024-019-0333-5. [ DOI ] [ Google Scholar ] [8]. Rombach R, Blattmann A, Lorenz D, Esser P, and Ommer B, “High-Resolution Image Synthesis with Latent Diffusion Models,” in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2022. doi: 10.1109/CVPR52688.2022.01042. [ DOI ] [ Google Scholar ] [9]. Puglisi L, Alexander DC, and Ravì D, “Enhancing Spatiotemporal Disease Progression Models via Latent Diffusion and Prior Knowledge,” May 2024, [Online]. Available: http://arxiv.org/abs/2405.03328 [ Google Scholar ] [10]. Pombo G et al. , “Equitable modelling of brain imaging by counterfactual augmentation with morphologically constrained 3D deep generative models,” Med Image Anal, vol. 84, 2023, doi: 10.1016/j.media.2022.102723. [ DOI ] [ Google Scholar ] [11]. Ravi D, Alexander DC, and Oxtoby NP, “Degenerative Adversarial NeuroImage Nets: Generating Images that Mimic Disease Progression,” in Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 2019. doi: 10.1007/978-3-030-32248-9_19. [ DOI ] [ Google Scholar ] [12]. Ho J, Jain A, and Abbeel P, “Denoising diffusion probabilistic models,” in Advances in Neural Information Processing Systems, 2020. [ Google Scholar ] [13]. Zhang L, Rao A, and Agrawala M, “Adding Conditional Control to Text-to-Image Diffusion Models,” in Proceedings of the IEEE International Conference on Computer Vision, 2023. doi: 10.1109/ICCV51070.2023.00355. [ DOI ] [ Google Scholar ] [14]. Koval I et al. , “AD Course Map charts Alzheimer’s disease progression,” Sci Rep, vol. 11, no. 1, 2021, doi: 10.1038/s41598-021-87434-1. [ DOI ] [ Google Scholar ] [15]. Schiratti J-B, Colliot O, and Durrleman S, “A Bayesian Mixed-Effects Model to Learn Trajectories of Changes from Repeated Manifold-Valued Observations,” 2017. [Online]. Available: http://jmlr.org/papers/v18/17-197.html . [16]. Jack CR et al. , “The Alzheimer’s Disease Neuroimaging Initiative (ADNI): MRI methods,” 2008. doi: 10.1002/jmri.21049. [ DOI ] [ Google Scholar ] [17]. Ellis KA et al. , “The Australian Imaging, Biomarkers and Lifestyle (AIBL) study of aging: Methodology and baseline characteristics of 1112 individuals recruited for a longitudinal study of Alzheimer’s disease,” Int Psychogeriatr, vol. 21, no. 4, 2009, doi: 10.1017/S1041610209009405. [ DOI ] [ Google Scholar ] [18]. LaMontagne PJ et al. , “OASIS-3: Longitudinal Neuroimaging, Clinical, and Cognitive Dataset for Normal Aging and Alzheimer Disease,” medRxiv, 2019. [ Google Scholar ] [19]. Yoon JS, Zhang C, Il Suk H, Guo J, and Li X, “SADM: Sequence-Aware Diffusion Model for Longitudinal Medical Image Generation,” in Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 2023. doi: 10.1007/978-3-031-34048-2_30. [ DOI ] [ Google Scholar ] [20]. Ferrucci L, “The baltimore longitudinal study of aging (BLSA): A 50-year-long journey and plans for the future,” 2008. doi: 10.1093/gerona/63.12.1416. [ DOI ] [ Google Scholar ] [21]. Shock N, Greulich R, Aremberg D, Costa P, Lakatta E, and Tobin J, Normal Human Aging: The Baltimore Longitudinal Study of Aging. Washington, DC: National Institutes of Health, 1984. [ Google Scholar ] [22]. Puglisi Lemuel, “Turboprep.” Accessed: Jul. 25, 2024. [Online]. Available: https://github.com/LemuelPuglisi/turboprep [23]. Tustison NJ et al. , “N4ITK: Improved N3 bias correction,” IEEE Trans Med Imaging, vol. 29, no. 6, 2010, doi: 10.1109/TMI.2010.2046908. [ DOI ] [ Google Scholar ] [24]. Hoopes A, Mora JS, Dalca AV, Fischl B, and Hoffmann M, “SynthStrip: skull-stripping for any brain image,” Neuroimage, vol. 260, 2022, doi: 10.1016/j.neuroimage.2022.119474. [ DOI ] [ Google Scholar ] [25]. Avants BB, Epstein CL, Grossman M, and Gee JC, “Symmetric diffeomorphic image registration with cross-correlation: Evaluating automated labeling of elderly and neurodegenerative brain,” Med Image Anal, vol. 12, no. 1, 2008, doi: 10.1016/j.media.2007.06.004. [ DOI ] [ Google Scholar ] [26]. Billot B et al. , “SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining,” Med Image Anal, vol. 86, 2023, doi: 10.1016/j.media.2023.102789. [ DOI ] [ Google Scholar ] [27]. Shinohara RT et al. , “Statistical normalization techniques for magnetic resonance imaging,” Neuroimage Clin, vol. 6, 2014, doi: 10.1016/j.nicl.2014.08.008. [ DOI ] [ Google Scholar ] [28]. Kao YH, Chou MC, Chen CH, and Yang YH, “White matter changes in patients with alzheimer’s disease and associated factors,” J Clin Med, vol. 8, no. 2, 2019, doi: 10.3390/jcm8020167. [ DOI ] [ Google Scholar ] [29]. Apostolova LG et al. , “Hippocampal atrophy and ventricular enlargement in normal aging, mild cognitive impairment (MCI), and Alzheimer disease,” Alzheimer Dis Assoc Disord, vol. 26, no. 1, 2012, doi: 10.1097/WAD.0b013e3182163b62. [ DOI ] [ Google Scholar ] ACTIONS View on publisher site PDF (1.2 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top