AD-DAE: Unsupervised Modeling of Longitudinal Alzheimer's Disease Progression with Diffusion Auto-Encoder

TL;DR

The paper tackles unsupervised modeling of longitudinal Alzheimer's disease progression from MRI by introducing AD-DAE, a diffusion auto-encoder that enables controlled latent shifts conditioned on progression attributes. By constraining the latent shift to a subspace and coupling it with a consistency module, AD-DAE generates anatomically plausible follow-up images from baselines while disentangling disease progression from subject identity. The approach yields improvements in image fidelity and volumetric progression accuracy, demonstrates robust cross-dataset generalization, and provides evidence of latent disentanglement through latent-swapping and manifold analyses. These capabilities support unsupervised, clinically meaningful trajectory generation and potential data augmentation for downstream disease classification. Overall, AD-DAE advances diffusion-based progression modeling by delivering controllable, unsupervised latent transitions aligned with anatomical changes observed in Alzheimer’s progression, across datasets and disease stages.

Abstract

Generative modeling frameworks have emerged as an effective approach to capture high-dimensional image distributions from large datasets without requiring domain-specific knowledge, a capability essential for longitudinal disease progression modeling. Recent generative modeling approaches have attempted to capture progression by mapping images into a latent representational space and then controlling and guiding the representations to generate follow-up images from a baseline image. However, existing approaches impose constraints on distribution learning, leading to latent spaces with limited controllability to generate follow-up images without explicit supervision from subject-specific longitudinal images. In order to enable controlled movements in the latent representational space and generate progression images from a baseline image in an unsupervised manner, we introduce a conditionable Diffusion Auto-encoder framework. The explicit encoding mechanism of image-diffusion auto-encoders forms a compact latent space capturing high-level semantics, providing means to disentangle information relevant for progression. Our approach leverages this latent space to condition and apply controlled shifts to baseline representations for generating follow-up. Controllability is induced by restricting these shifts to a subspace, thereby isolating progression-related factors from subject identity-preserving components. The shifts are implicitly guided by correlating with progression attributes, without requiring subject-specific longitudinal supervision. We validate the generations through image quality metrics, volumetric progression analysis, and downstream classification in Alzheimer's disease datasets from two different sources and disease categories. This demonstrates the effectiveness of our approach for Alzheimer's progression modeling and longitudinal image generation.

Figures (11)

• Figure 1: From left to right, the training and inference strategy is described. Training Module: From left to right, the baseline image ($x_b$), Gaussian noise ($x_T$), time embeddings ($t$) and the progression attributes ($v_a,v_d$) are processed through the encoder ($\mathcal{E}$), latent shift module ($\mathcal{A}$) and the decoder ($\mathcal{D}$). The latent vector ($z_b$) from $\mathcal{E}$ gets shifted by $z^{'}$ incorporating progression factors and generating follow-up image ($\hat{x}_f$). The regression component $\mathcal{R}$ processes $x_b$, $\hat{x}_f$, and their residual to estimate the progression attributes, while optimizing $\mathcal{A}$. Inference Module: Image $x_b$, noise, and progression attributes are processed $T_s$ times to generate $\hat{x}_f$ with latent shift integration.
• Figure 2: From left to right, columns (a)–(g) show predicted follow-up images ($\hat{x}_f$) with their error maps ($|\hat{x}_f - x_f|$) for AD-DAE and baseline methods, and column (h) shows the ground truth follow-up images ($x_f$). Results are shown for a subject from the MCI category (Test Set) with a baseline age of 75 years and a four-year age span (76–79 years). The error maps from AD-DAE exhibit relatively lower errors compared to other models.
• Figure 3: From left to right, columns (a)–(g) show the generated baseline ($\hat{x}_b$), follow-up ($\hat{x}_f$), and progression-related changes through the difference ($|x_b - \hat{x}_f|$) for AD-DAE and comparative methods; column (h) shows the ground-truth baseline and follow-up. Results are from a subject of MCI category (Test Set), with a baseline age of 75 years and a four-year progression span (76–79 years). The difference maps indicate that AD-DAE captures the progression hierarchy relatively closer to the ground-truth progression changes.
• Figure 4: Top to bottom: From Test Set images and error maps for CN, MCI, and AD cohorts. Left to right: (a) Ground truth baselines (age 82), (b) Ground truth follow-ups (age 84), (c)–(i) Generated follow-up images and error maps (relative to ground truth follow-up) for AD-DAE and baseline methods. Overall, AD-DAE shows relatively lower errors in comparison to the other methods.
• Figure 5: Left to right: Performance comparison of AD-DAE on Cross-Data Setup with baseline methods in terms of (a) PSNR, (b) SSIM, and (c) MSE. CN is shown in green and MCI & AD in violet, represented with their mean and first-quartile values. Across all metrics, AD-DAE demonstrates relatively better performance.