Abstract
White matter hyperintensities (WMHs) are commonly scored using the ordinal Fazekas scale, but manual assessment is subjective and laborious to obtain. We propose RBT-Faz, a direct 3D fluid-attenuated inversion recovery (FLAIR) MRI framework for subject-level periventricular hyperintensity (PVH) and deep white matter hyperintensity (DWMH) Fazekas scoring. RBT-Faz uses a 3D Swin-UNETR encoder and introduces a final-layer Firewall Attention Block, where learnable register tokens aggregate global WMH severity context from localized patch tokens through a register bottleneck. This design preserves lesion-sensitive spatial features while forming compact subject-level severity representations. To improve ordinal consistency and clinical grounding, the visual embedding is aligned with ClinicalBERT-encoded ground-truth Fazekas grade descriptors, and a lesion-aware auxiliary head predicts WMH burden, periventricular/deep spatial distribution, and lesion count during training only. Experiments on CCNA subjects showed that RBT-Faz outperformed volumetric classifiers and segmentation-based rule grading, achieving Macro-F1/QWK of 0.786/0.886 for PVH and 0.825/0.914 for DWMH. Grad-CAM and descriptor analyses further supported clinically meaningful model behavior. The anonymized code repository is available at: \url{https://github.com/IAMLAB-Ryerson/RBT-Faz}
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BrainWorks_025.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=%7EApril_Khademi1
BibTex
@InProceedings{MazFar_RBTFaz_MICCAISAT2026,
author = { Mazandarani, Farzan Niknejad AND Tsang, Kai La Jennifer AND Simard, Nicholas AND Aktar, Mumu AND Khademi, April},
title = { { RBT-Faz: Register-Bottleneck 3D Swin-UNETR with Clinical Text Alignment for Ordinal Fazekas Grading from FLAIR MRI } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17254},
month = {pending},
page = {pending}
}