CMIF: Cross-Modality Interactive Fusion for Diffusion-Based MRI Synthesis and Its Impact on Brain Tumor Segmentation
Sensors · 20.11.2026 · 26 (22) · 7347
DOI: 10.3390/s26227347 ↗Kurz gefasst
Fachartikel in der Zeitschrift Sensors, online erschienen am 20.11.2026. Verfasst von Jiu Bak und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Missing MRI modalities can impair brain tumor segmentation, but synthesizing a missing modality does not necessarily improve downstream performance. We propose a conditional wavelet diffusion framework incorporating Cross-Modality Interactive Fusion (CMIF) and evaluate its segmentation utility. The diffusion model was trained on BraTS 2021, and the segmentation model trained on BraTS 2018 was evaluated on BraTS-Africa 2024 without additional training. CMIF was compared with cWDM using SSIM, PSNR, and downstream Dice scores under missing, generated, and fully acquired input conditions. CMIF achieved higher SSIM and PSNR for T1ce and FLAIR, but improvements in the performance of the segmentation model varied depending on the modalities, model, and tumor region. For generated T1ce, model-averaged WT Dice score increased from 70.57 with cWDM to 71.43 with CMIF, whereas ET Dice score decreased from 24.92 to 23.76. Spatial error analyses identified both reduced False Positive predictions and increased tumor omissions. These findings show that image similarity alone is insufficient to assess the utility of generated MRI and support evaluating modality completion through both regional segmentation performance and spatial error patterns.
Zitieren
APA
Bak, J., Kim, G. H., & Choi, H. (2026). CMIF: Cross-Modality Interactive Fusion for Diffusion-Based MRI Synthesis and Its Impact on Brain Tumor Segmentation. Sensors, 26(22), 7347. https://doi.org/10.3390/s26227347
Vancouver
Bak J, Kim GH, Choi H. CMIF: Cross-Modality Interactive Fusion for Diffusion-Based MRI Synthesis and Its Impact on Brain Tumor Segmentation. Sensors. 2026;26(22):7347. doi:10.3390/s26227347
BibTeX
@article{Bak2026,
author = {Bak, Jiu and Kim, Gi Hoon and Choi, Hyunguk},
title = {CMIF: Cross-Modality Interactive Fusion for Diffusion-Based MRI Synthesis and Its Impact on Brain Tumor Segmentation},
journal = {Sensors},
year = {2026},
volume = {26},
number = {22},
pages = {7347},
doi = {10.3390/s26227347}
}
EUVIMED