A technical assessment of latent diffusion for Alzheimer's disease progression

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.