東北大学 大学院情報科学研究科 情報基礎科学専攻 計算機構論分野
(東北大学 工学部 電気情報物理工学科 情報工学コース)
青木・伊藤(康)研究室

SQET-MoE: Two-Stage Brain Age Estimation Using Squeeze-and-Excitation Transformer and Mixture-of-Experts

Rizuki Oura (Tohoku University) , Koichi Ito (Tohoku University) , Takafumi Aoki (Tohoku University)
International Symposium on Biomedical Imaging, April 2026.
Abstract

Brain age estimation is crucial for identifying brain disorders and developing biomarkers. While deep learning methods have been proposed for high-accuracy age estimation from T1-weighted images, further improvements are still needed. This paper proposes a novel brain age estimation method called SQET-MoE, utilizing a two-stage estimation framework. In the 1st stage, the Squeeze-and-Excitation Transformer (SQET), which fuses the benefits of CNNs and Transformers, performs coarse age estimation and feature extraction. The 2nd stage employs a Mixture-of-Experts (MoE) module, which uses SQET's output to estimate and correct the residual associated with systematic errors and complex non-linearity. This task division stabilizes training and maximizes estimation accuracy. Through a set of experiments using large-scale datasets, we demonstrate that the proposed SQET-MoE achieves the highest estimation accuracy compared to conventional methods.

キーワード
戻る