Identity Preserving Latent Diffusion for Brain Aging Modeling

TL;DR

This work introduces IP-LDM, an age- and identity-conditioned latent diffusion model designed for longitudinal brain aging synthesis that preserves intra-subject identity. By combining an age encoder, a triplet-contrastive identity representation learner, and an identity-controlled latent diffusion transformation, the method enables continuous and realistic age progression from a single source brain image. Empirical results on elderly (OASIS-3) and infant (IBIS/Baby Brain) datasets show superior identity preservation and image quality, with quantitative gains in SSIM, PSNR, FID, KID, RMSE, and ARI over strong baselines. The approach offers a practical tool for studying brain aging dynamics and could impact neurodevelopmental and neurodegenerative research, though it is presently limited by dataset size and 2D processing, with future work aimed at 3D extensions and broader generalization.

Abstract

Structural and appearance changes in brain imaging over time are crucial indicators of neurodevelopment and neurodegeneration. The rapid advancement of large-scale generative models provides a promising backbone for modeling these complex global and local changes in brain images, such as transforming the age of a source image to a target age. However, current generative models, typically trained on independently and identically distributed (i.i.d.) data, may struggle to maintain intra-subject spatiotemporal consistency during transformations. We propose the Identity-Preserving Longitudinal Diffusion Model (IP-LDM), designed to accurately transform brain ages while preserving subject identity. Our approach involves first extracting the identity representation from the source image. Then, conditioned on the target age, the latent diffusion model learns to generate the age-transformed target image. To ensure consistency within the same subject over time, we regularize the identity representation using a triplet contrastive formulation. Our experiments on both elderly and infant brain datasets demonstrate that our model outperforms existing conditional generative models, producing realistic age transformations while preserving intra-subject identity.

Figures (10)

• Figure 1: Overview of proposed longitudinal diffusion model. a) shows the overall architecture of IP-LDM, consisting of an age and identity conditioned latent diffusion, wherein $\mathcal{E}$ and $\mathcal{D}$ are the image encoder and decoder, respectively. b) illustrates the details of the identity representation learning along with the learned feature distributions (different color indicates different subjects). c) depicts the forward and backward process in the latent manipulation module.
• Figure 2: Visualization of baseline comparison on the Baby Brain and OASIS-3 datasets. The brain images are generated by different models based on the source images, which are expected to align with the target images. In the first row, the source and target are similar in age, resulting in subtle yet discernible longitudinal changes. Our method preserves the identity more effectively compared to other non-identity-preserving baselines.
• Figure 3: Qualitative visualization for brain age transformation. Each row represents brain images generated at different ages. The first row showcases the results from the proposed method IP-LDM. Subsequent rows display results from other methods, including InstructPix2Pix, DAE, and cGAN. Each column represents brain images generated at specific ages, ranging from age 35 to age 85.
• Figure 4: Inputs and reconstructions of VAE trained on Baby Brain. The first and third rows display the input baby brain MR images. The second and fourth rows show the reconstructed baby brain images.
• Figure 5: Inputs and reconstructions of VAE trained on OASIS-3. The first and third rows display the input OASIS-3 brain MR images. The second and fourth rows show the reconstructed OASIS-3 brain images.