Abstract
Differential diagnosis of neurodegenerative diseases using 3D T1w MRI remains a challenge, particularly when distinguishing between clinically similar dementia subtypes. While deep learning has achieved high performance on binary screening (e.g., controls vs. Alzheimer’s Disease), standard architectures often struggle with the class imbalance and subtle anatomical variations, leading to a critical trade-off between global specificity and focal sensitivity. In this work, we propose a hybrid Dual-Stream framework using a Hierarchical Differential MedViT-3D. First, a fine-scale stream adapts the MedViTV2 architecture to 3D and adds recent Differential Attention and Rotary Positional Embeddings (RoPE) mechanisms to capture focal anomalies that standard attention misses. Second, to counter the “vanishing specificity” of such high-sensitivity models, we fuse this focal stream with global volumetric priors and introduce a hierarchical loss strategy grouping pathologies. We evaluate our method on a challenging 7-class diagnostic task using over 4,000 scans aggregated from 16 datasets. Our experiments suggest that this dual-stream hierarchical specialization improves the empirical trade-off between rare-variant sensitivity and CN specificity, achieving the highest Top-2 balanced accuracy among the evaluated methods on the held-out OOD set. By explicitly modeling the balance between global context and local saliency, our approach provides a clinically motivated framework for computer-aided diagnosis in complex settings. Code is available here: https://github.com/EloiNavet/Hierarchical-Differential-MedViT-3D.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BrainWorks_005.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/BrainWorks_005_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/profile?id=~Eloi_Navet1
BibTex
@InProceedings{NavÉlo_Hierarchical_MICCAISAT2026,
author = { Navet, Éloi AND Giraud, Rémi AND Mansencal, Boris AND Coupé, Pierrick},
title = { { Hierarchical Differential MedViT-3D: Specialized Hybrid Transformer for Multiclass Diagnosis of Neurodegenerative Diseases } },
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}
}
