Dual-path mixture of experts for Alzheimer's diagnosis using volume and entropy MRI features

Alzheimer's disease (AD) is a progressive neurodegenerative disorder marked by brain atrophy, observable via MRI. Volumetric measures like hippocampal volume are widely used but often miss early microstructural changes. Texture features such as intensity entropy capture tissue heterogeneity and may improve early diagnosis. Combining these orthogonal modalities could yield better biomarkers.

We propose a Dual-Path Mixture of Experts (DP-MoE) model that integrates volumetric and entropy-based MRI features through two expert subnetworks and a self-attention gating mechanism. A residual fusion branch preserves raw input distributions, enhancing flexibility and expressiveness. The code is available at https://github.com/syoonni/DP-MOE.

The model was evaluated on three binary classification tasks (CN vs. AD, CN vs. MCI, MCI vs. AD) using T1-weighted and entropy-derived MRI. For CN vs. AD, it achieved an accuracy of 93.02% (0.9302 ± 0.0109) and an AUC of 0.975 ± 0.004, outperforming both the baseline MoE and logistic fusion models. Ablation studies confirmed the importance of both attention and residual components.

Prior methods used concatenation or fixed weighting, neglecting modality-specific characteristics. Volume and entropy provide orthogonal, complementary information. While Mixture of Experts is common in other fields, it remains underused in AD imaging. Our adaptive fusion approach improved both interpretability and accuracy, especially in early-stage discrimination.

DP-MoE offers an effective, interpretable framework for fusing heterogeneous MRI biomarkers and holds promise for deployment in early Alzheimer's diagnosis workflows.