MISFIT: Modality Inference via Style Fusion and Invertible Translation for Cross-Modality Synthesis of 3D MRI Volumes
Banerjee, Mohor Ji, Qiankun Hashmi, Sarim Elsayed, Abdelrahman Tikhonov, Daniil Scatolin, Matheus Ferracciu Jaheen, Ahmed Kim, Damir Wang, Hu Salem, Mostafa
Banerjee, Mohor
Ji, Qiankun
Hashmi, Sarim
Elsayed, Abdelrahman
Tikhonov, Daniil
Scatolin, Matheus Ferracciu
Jaheen, Ahmed
Kim, Damir
Wang, Hu
Salem, Mostafa
Supervisor
Department
Computer Vision
Embargo End Date
Type
Conference proceeding
Date
License
Language
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Robust synthesis of missing MRI modalities is essential for enabling the deployment of brain tumor segmentation pipelines in real-world clinical settings, where complete multi-modal scans are possibly unavailable due to acquisition constraints, scanner limitations, or patient-specific factors. This critical limitation is the focus of the Brain MR Image Synthesis Challenge (BraSyn), Task 8 of the BraTS-Lighthouse 2025 Challenge, which aims to foster generalizable and clinically deployable methods for synthesizing missing MRI modalities from available ones. To tackle this, we propose MISFIT – Modality Inference via Style Fusion and Invertible Translation – a novel two-stage generative framework for 3D brain MRI synthesis that operates entirely in the wavelet domain. MISFIT enhances the conditional wavelet diffusion paradigm by incorporating a Channel Attention Feature Fusion (CAFF) module for modality-aware representation learning and a Multi-Switchable SPADE block for style-conditioned generation. In Stage I, fused wavelet representations of the input modalities are used to generate a coarse approximation of the target modality. In Stage II, a denoising diffusion model - conditioned on both Stage I outputs and the original wavelet subbands - progressively reconstructs the final high-fidelity target volume. Evaluated on the BraSyn 2025 validation set, MISFIT achieves a PSNR of 18.65 dB and SSIM of 0.8041 in the randomized setting, demonstrating the viability of wavelet-based diffusion with adaptive fusion for medical image reconstruction. https://github.com/mohrsalt/MISFIT
Citation
M. Banerjee, Q. Ji, S. Hashmi, A. Elsayed, D. Tikhonov, M.F. Scatolin , et al., "MISFIT: Modality Inference via Style Fusion and Invertible Translation for Cross-Modality Synthesis of 3D MRI Volumes," 2026, pp. 42-53.
Source
Lecture Notes in Computer Science
Conference
International Conference on Medical Image Computing and Computer-Assisted Intervention
Keywords
46 Information and Computing Sciences, 4603 Computer Vision and Multimedia Computation
Subjects
Source
International Conference on Medical Image Computing and Computer-Assisted Intervention
Publisher
Springer Nature