Abstract
Data scarcity and class imbalance in brain tumor datasets remain critical bottlenecks for training robust diagnostic AI models. Synthesizing realistic tumors across multiple MRI modalities is challenging because sequences differ in sensitivity to pathological subregions, modality quality varies per subject, and strict inter-modal consistency must be maintained. We propose a subject-specific multimodal latent diffusion framework for high-fidelity brain tumor synthesis via Disease- Aware Modality AttentioN (DAMAN). Separate modality-specific 3D VQ-GANs encode each MRI contrast into a compact latent space; our framework dynamically computes global quality-aware gating weights and voxel-level lesion-specific spatial attention to adaptively fuse complementary cross-modal features for each patient. A hard-inpainting loop at every reverse diffusion step enforces strict background preservation. Evaluated on the LUMIERE longitudinal glioblastoma dataset with four MRI contrasts under a patient-level train/test split, our framework achieves PSNR of 20.27 dB and SSIM of 0.919 on the target modality (T1-post), outperforming single-modal baselines. An independent, harmonizationfree, whole-image evaluation across all four modalities further confirms these gains and shows that DAMAN reaches a 100% valid generation rate, compared with 50–56% for the baselines.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/DGM4MICCAI_012.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/profile?id=~Kyuwon_Park1
BibTex
@InProceedings{ParKyu_PatientAdaptive_MICCAISAT2026,
author = { Park, Kyuwon AND Liu, Xiaofeng AND Xing, Fangxu AND Shih, Helen A. AND El Fakhri, Georges AND Hwang, Jae Youn AND Woo, Jonghye},
title = { { Patient-Adaptive Modality Attention for Multimodal Brain Tumor Synthesis via Latent Diffusion } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17258},
month = {pending},
page = {pending}
}