ADP-DiT: Text-Guided Diffusion Transformer for Brain Image Generation in Alzheimer's Disease Progression

Juneyong Lee; Geonwoo Baek; Ikbeom Jang

cs.CVApr 15, 2026

ADP-DiT: Text-Guided Diffusion Transformer for Brain Image Generation in Alzheimer's Disease Progression

Authors: Juneyong Lee, Geonwoo Baek, Ikbeom Jang

Organizations: Department of Computer Science and Engineering, Hankuk University of Foreign Studies, Seoul, Republic of Korea

Abstract

Alzheimer's disease (AD) progresses heterogeneously across individuals, motivating subject-specific synthesis of follow-up magnetic resonance imaging (MRI) to support progression assessment. While Diffusion Transformers (DiT), an emerging transformer-based diffusion model, offer a scalable backbone for image synthesis, longitudinal AD MRI generation with clinically interpretable control over follow-up time and participant metadata remains underexplored. We present ADP-DiT, an interval-aware, clinically text-conditioned diffusion transformer for longitudinal AD MRI synthesis. ADP-DiT encodes follow-up interval together with multi-domain demographic, diagnostic (CN/MCI/AD), and neuropsychological information as a natural-language prompt, enabling time-specific control beyond coarse diagnostic stages. To inject this conditioning effectively, we use dual text encoders-OpenCLIP for vision-language alignment and T5 for richer clinical-language understanding. Their embeddings are fused into DiT through cross-attention for fine-grained guidance and adaptive layer normalization for global modulation. We further enhance anatomical fidelity by applying rotary positional embeddings to image tokens and performing diffusion in a pre-trained SDXL-VAE latent space to enable efficient high-resolution reconstruction. On 3,321 longitudinal 3T T1-weighted scans from 712 participants (259,038 image slices), ADP-DiT achieves SSIM 0.8739 and PSNR 29.32 dB, improving over a DiT baseline by +0.1087 SSIM and +6.08 dB PSNR while capturing progression-related changes such as ventricular enlargement and shrinking hippocampus. These results suggest that integrating comprehensive, subject-specific clinical conditions with architectures can improve longitudinal AD MRI synthesis.

Figures & tables

AD progression group	CN to CN	CN to MCI	MCI to MCI	MCI to AD	AD to AD
Number of participants	160	27	434	117	106
Number of scans	755	43	1909	234	380
Visits per participant	5.0 (2–6)	4.7 (2–6)	4.9 (1–9)	4.7 (2–8)	3.7 (2–7)
Interval from baseline (month)	13.0 $\pm$ 15.0	20.6 $\pm$ 27.0	13.1 $\pm$ 15.4	17.6 $\pm$ 21.1	6.5 $\pm$ 10.0
Age (year)	72.7 $\pm$ 5.5	76.2 $\pm$ 7.0	71.0 $\pm$ 7.4	72.3 $\pm$ 7.3	75.4 $\pm$ 7.5
Sex (female/male)	61 / 55	10 / 15	119 / 158	41 / 66	37 / 57

Table 1: Demographic, cognitive, and clinical characteristics of participants by AD progression group.

Figure 1: The overview of the proposed ADP-DiT framework. The model conditions on multiple text-encoders text embeddings and structured metadata to guide the denoising process in the latent space.

Models	SSIM $\uparrow$	PSNR $\uparrow$	MSE $\downarrow$
Baseline (Out = In)	0.8739 $\pm$ 0.0640	26.14 $\pm$ 2.654 ∗∗	0.0030 $\pm$ 0.0024 ∗∗
Diffusion Transformer [ 1 ]	0.7652 $\pm$ 0.0599 ∗∗	23.24 $\pm$ 1.840 ∗∗	0.0052 $\pm$ 0.0024 ∗∗
SwinUNETR-V2 [ 11 ]	0.8535 $\pm$ 0.0686 ∗∗	28.37 $\pm$ 3.108 ∗∗	0.0019 $\pm$ 0.0019
FCDiffusion [ 10 ]	0.5474 $\pm$ 0.1394 ∗∗	20.55 $\pm$ 2.562 ∗∗	0.0106 $\pm$ 0.0080 ∗∗
Diffusion-CLIP [ 8 ]	0.6556 $\pm$ 0.1124 ∗∗	24.33 $\pm$ 3.707 ∗∗	0.0051 $\pm$ 0.0041 ∗∗
VQGAN-CLIP [ 6 ]	0.7463 $\pm$ 0.0594 ∗∗	22.97 $\pm$ 1.816 ∗∗	0.0055 $\pm$ 0.0024 ∗∗

Table 2: Quantitative comparison of ADP-DiT against state-of-the-art models. Bold indicates the best performance among generative models, and underline indicates the second best. Models below the dashed line are sorted by SSIM.

Progression	Interval (Mo.)	N	Scan	SSIM $\uparrow$	PSNR $\uparrow$	MSE $\downarrow$
CN $\to$ CN	8.50 $\pm$ 7.02	30	171	0.8695 $\pm$ 0.0804	29.09 $\pm$ 6.739	0.0025 $\pm$ 0.0029
CN $\to$ MCI	13.31 $\pm$ 9.19	5	10	0.8131 $\pm$ 0.0233	24.58 $\pm$ 1.256	0.0036 $\pm$ 0.0010
MCI $\to$ MCI	7.49 $\pm$ 5.59	81	432	0.8779 $\pm$ 0.0727	29.60 $\pm$ 6.850	0.0023 $\pm$ 0.0021
MCI $\to$ AD	11.08 $\pm$ 6.60	21	53	0.8227 $\pm$ 0.0685	24.73 $\pm$ 2.697	0.0041 $\pm$ 0.0034
AD $\to$ AD	7.43 $\pm$ 6.09	20	86	0.8895 $\pm$ 0.0761	30.71 $\pm$ 7.236	0.0020 $\pm$ 0.0018

Table 3: Performance of ADP-DiT across progression groups.

Interval (Mo.)	Models	SSIM $\uparrow$	PSNR $\uparrow$	MSE $\downarrow$
$0\leq\Delta t<12$ #subj: 151 #slice: 35334	Diffusion Transformer [ 1 ]	0.7748 $\pm$ 0.0532	23.51 $\pm$ 1.788	0.0048 $\pm$ 0.0021
	SwinUNETR-V2 [ 11 ]	0.8663 $\pm$ 0.0632	29.29 $\pm$ 2.954	0.0015 $\pm$ 0.0014
	FCDiffusion [ 10 ]	0.5634 $\pm$ 0.1392	20.82 $\pm$ 2.578	0.0100 $\pm$ 0.0076
	Diffusion-CLIP [ 8 ]	0.6651 $\pm$ 0.1168	25.14 $\pm$ 3.973	0.0044 $\pm$ 0.0037
	VQGAN-CLIP [ 6 ]	0.7553 $\pm$ 0.0523	23.21 $\pm$ 1.774	0.0052 $\pm$ 0.0021
	Stable Diffusion 2.1 [ 9 ]	0.7172 $\pm$ 0.1151	24.41 $\pm$ 2.706	0.0044 $\pm$ 0.0033

Table 4: Benchmark results comparing ADP-DiT with state-of-the-art methods across specific time intervals. Bold indicates the best performance among the compared methods, and underline indicates the second best. Here,

\Delta t

denotes the time elapsed (in months) between the baseline input and the target follow-up scan.

Figure 2: Visual comparison of ADP-DiT against benchmark models. The absolute error maps highlight the voxel-wise progression predicted by each model relative to the ground truth, demonstrating the superior accuracy of ADP-DiT.

Figure 3: Results of ADP-DiT for AD progression. The absolute error map visualizes voxel-wise discrepancies between the generated output and the ground truth.

Explore similar work

Structural MRI Synthesis for Alzheimer's Disease via Conditional Diffusion on Anatomical Masks
Jun 16, 2026Muge Zhang, Muhammad Ali Khaliq, Jamal Alsakran +2Synthetic Data Augmentation Alzheimer's Disease
Fairleigh Dickinson University, Vancouver, BC, Canada · University of Colorado at Colorado Springs, Colorado Springs, CO, USA
Adaptive Clinical-Aware Latent Diffusion for Multimodal Brain Image Generation and Missing Modality Imputation
Mar 10, 2026Rong Zhou, Houliang Zhou, Yao Su +4Medical Image Synthesis Missing-Modality Learning
Department of Computer Science and Engineering, Lehigh University, PA, USA · Worcester Polytechnic Institute, MA, USA · Department of Psychiatry and Behavioral Sciences, Stanford University School of Medicine, CA, USA +3
AD-DAE: Alzheimer's Disease Progression Modeling with Unpaired Longitudinal MRI using Diffusion Auto-Encoders
Nov 8, 2025Ayantika Das, Arunima Sarkar, Keerthi Ram +1Medical Image Synthesis Alzheimer's Disease Progression Prediction
Department of Electrical Engineering, Indian Institute of Technology Madras (IITM), Chennai 600036, India · Sudha Gopalakrishnan Brain Centre (SGBC), IITM, Chennai 600036, India

Jun 16, 2026J/K move · Enter open · S save

Structural MRI Synthesis for Alzheimer's Disease via Conditional Diffusion on Anatomical Masks

Muge Zhang, Muhammad Ali Khaliq, Jamal Alsakran, Byeong Kil Lee+1

Fairleigh Dickinson University, Vancouver, BC, Canada · University of Colorado at Colorado Springs, Colorado Springs, CO, USA

Synthetic Data Augmentation Alzheimer's Disease Medical Image Synthesis Conditional Diffusion Models

Open paper AskPDF arXiv

Recent advances in generative machine learning models have significantly improved medical imaging, offering promising solutions for data augmentation, privacy preservation, and improved model generalization. However, synthesizing high-quality structural MRI data for Alzheimer's Disease (AD) remains challenging due to the subtle, region-specific, and progressive anatomical changes associated with neurodegeneration. In this paper, we extend the Med-DDPM conditional diffusion model -- originally designed for brain tumor synthesis -- to generate 3D structural MRIs specifically tailored to AD. We adopted Med-DDPM due to its established stability and structural fidelity compared to other generative models, which makes it particularly suitable for capturing the subtle anatomical changes characteristic of AD. Our approach conditions the diffusion process on anatomical segmentation masks derived from the ADNI dataset, incorporating key AD-relevant brain structures into the generation process. We systematically evaluate the quality and utility of the synthetic images by training segmentation models on real, synthetic, and hybrid (mixed) datasets. Experimental results demonstrate that segmentation models trained exclusively on synthetic data achieve comparable Dice scores (0.6532) to those trained on real data (0.6513), while exhibiting significantly enhanced recall. Notably, models trained on hybrid datasets (mixing real and synthetic images) outperform both real and synthetic-only baselines, achieving a Dice score of 0.7244. These findings underscore the successful use of conditional diffusion models for generating anatomically accurate, AD-specific synthetic MRIs, and highlight their potential for enhancing training data availability, improving diagnostic accuracy, and promoting research reproducibility in neuroimaging studies.

Abstract

Explore similar work

Structural MRI Synthesis for Alzheimer's Disease via Conditional Diffusion on Anatomical Masks

Adaptive Clinical-Aware Latent Diffusion for Multimodal Brain Image Generation and Missing Modality Imputation

AD-DAE: Alzheimer's Disease Progression Modeling with Unpaired Longitudinal MRI using Diffusion Auto-Encoders