Align-cDAE: Attention-Aligned Conditional Diffusion Auto-encoder for Alzheimer's Disease Progression

Ayantika Das, Keerthi Ram, Mohanasankar Sivaprakasam

Keywords: Multi-modal Conditions, Conditional Alignment, Denoising Diffusion Model, Progression Modeling, Alzheimer’s Disease.

TL;DR: Attention-Aligned Conditional Diffusion Auto-encoder

Abstract: The integration of multi-modal conditioning with diffusion modeling approaches has been shown to be effective in image-to-image translation tasks. Existing mechanisms usually integrate conditioning information from other modalities by projecting it into a feature space compatible with the image representations. Although these strategies yield improved performance, they do not ensure that information from non-imaging conditioning modalities meaningfully aligns with image features and precisely modulates the generated outputs. In order to better incorporate information from other modalities, we propose a diffusion auto-encoder-based framework for disease progression modeling that explicitly focuses on conditional alignment. This alignment is introduced by constraining the attention between (i) the conditioning attributes and (ii) the feature representations of the model, to focus on the regions exhibiting progression-related changes. This constraint consequently shifts the attention of different layers of the model towards progression-specific regions, generating the required precise anatomical changes. Further, the diffusion auto-encoding-based formulation provides latent representations of images that are compact in nature and suitable for integration of conditions. We have experimentally validated the performance of our model by evaluating on Alzheimer's disease progression generation through various image-level metrics and volumetric assessments. These results demonstrate that enforcing conditional alignment within a diffusion auto-encoding framework leads to more anatomically precise modeling of Alzheimer's disease progression.

Primary Subject Area: Image Synthesis

Secondary Subject Area: Generative Models

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 326