Parameter-Efficient Mask-Conditioned Magnetic Resonance Image Synthesis for Segmentation Augmentation
Abstract When real training images for magnetic resonance imaging (MRI) segmentation are limited, existing masks can be reused to synthesize additional image appearances. However, paired image-translation methods may favor pixel fidelity over distributional similarity, while full diffusion-model fine-tuning is costly and generic conditioning does not explicitly encode mask geometry. To address these limitations, we propose ControlNet-guided Low-Rank Adaptation for MRI synthesis (\method), a mask-conditioned latent diffusion framework that adapts both the denoising and spatial-conditioning pathways. Only 2.96\% of the framework parameters are trainable in the BraTS configuration. We evaluate image quality and segmentation utility on BraTS2020 brain MRI and ACDC cardiac MRI across four fractions of real training slices. On BraTS T1 contrast-enhanced MRI, \method achieves a lower FID than Pix2Pix (55.65 versus 162.42), although Pix2Pix has higher paired pixel fidelity. With 10\% real training slices, adding synthetic pairs raises mean Dice from 0.7208 to 0.7401 and reduces HD95 from 48.43 to 45.34 pixels in a within-cohort, slice-level BraTS evaluation. On 50 held-out ACDC patients, a separately trained adapter raises mean Dice from 0.7661 to 0.8263 and reduces HD95 from 15.40 to 13.63 pixels at the same fraction. Gains diminish as real-image availability increases. These results support targeted use of parameter-efficient, mask-conditioned MRI synthesis when real-image coverage is low.
Authors
- Dongfang Chen (ORCID: https://orcid.org/0000-0002-7813-2295)
- xin 童鑫 Tong (ORCID: https://orcid.org/0009-0005-3386-0466)
- Liang 春辉 Chunhui
- Jianye An
Institutions
- Tianjin University of Commerce (CN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-12
- DOI
- https://doi.org/10.21203/rs.3.rs-11479149/v1
- Primary Topic
- Medical Image Segmentation Techniques
- Type
- preprint