Abstract
Diffusion models in medical imaging have emerged as important priors suited for numerous downstream analysis tasks. Training a separate 3D diffusion model from scratch for each MRI contrast is computationally expensive and difficult to scale. Pretrained models enable data- and time-efficient adaptation, but conventional fine-tuning requires maintaining a complete set of model weights per target contrast. We study Low-Rank Adaptation (LoRA) as a parameter-efficient mechanism for contrast specialization of pretrained 3D MRI diffusion models without modifying the backbone weights. Adapting a pretrained T_1-weighted model to T_2-weighted MRI, we analyze model behavior under different adapter ranks and training hyperparameters, evaluating anatomical distribution alignment with WASABI and perceptual similarity with FID. LoRA achieves anatomical alignment comparable to full fine-tuning (WASABI 2.10 vs. 2.29) while updating only 9.6% of model parameters, with modest but still competitive perceptual quality (FID 38.41 vs. 34.37). The gap between WASABI and FID indicates that low-rank updates recover global anatomical alignment more readily than fine-grained perceptual texture. The resulting adapters are lightweight and interchangeable: a rank-16 adapter requires 15.04 MB compared with 157.34 MB for a full fine-tuned checkpoint. The source code and model weights are made available at: https://github.com/piksl-research/t2w-lora-diffusion.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/EMA_022.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: Not Available
BibTex
@InProceedings{DewBla_LowRank_MICCAISAT2026,
author = { Dewey, Blake E. AND Remedios, Samuel W. AND Chaudhary, Muhammad F. A. AND Carass, Aaron AND Prince, Jerry L.},
title = { { Low-Rank Adaptation for Efficient Contrast Specialization of 3D MRI Diffusion Models } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17265},
month = {pending},
page = {pending}
}