REN: Anatomically-Informed Mixture-of-Experts for Interstitial Lung Disease Diagnosis

TL;DR

This work introduces Regional Expert Networks (REN), the first anatomically-informed Mixture-of-Experts framework for medical image classification, specifically ILD. By assigning seven region-specific experts to lung lobes and bilateral lungs and integrating radiomics with deep learning through multi-modal gating, REN achieves a substantial performance gain, achieving an average AUC of $0.8646\pm0.0467$, a $12.5\%$ improvement over the SwinUNETR baseline ($AUC=0.7685$, $p=0.031$), with lower-lobe AUCs around $0.88$–$0.90$. The approach emphasizes region-aware interpretability, demonstrating that learned gating can adaptively weight anatomically critical regions and that radiomics signals particularly enhance lower-lobe ILD detection. The methodology combines four stages—region extraction, regional expert training, gating strategies, and end-to-end MoE integration—yielding a scalable framework with potential applicability to other structured medical imaging tasks, while acknowledging limitations such as single-institution data and a focus on a specific ILD subtype for future validation and extension.

Abstract

Mixture-of-Experts (MoE) architectures have significantly contributed to scalable machine learning by enabling specialized subnetworks to tackle complex tasks efficiently. However, traditional MoE systems lack domain-specific constraints essential for medical imaging, where anatomical structure and regional disease heterogeneity strongly influence pathological patterns. Here, we introduce Regional Expert Networks (REN), the first anatomically-informed MoE framework tailored specifically for medical image classification. REN leverages anatomical priors to train seven specialized experts, each dedicated to distinct lung lobes and bilateral lung combinations, enabling precise modeling of region-specific pathological variations. Multi-modal gating mechanisms dynamically integrate radiomics biomarkers and deep learning (DL) features (CNN, ViT, Mamba) to weight expert contributions optimally. Applied to interstitial lung disease (ILD) classification, REN achieves consistently superior performance: the radiomics-guided ensemble reached an average AUC of 0.8646 +/- 0.0467, a +12.5 percent improvement over the SwinUNETR baseline (AUC 0.7685, p = 0.031). Region-specific experts further revealed that lower-lobe models achieved AUCs of 0.88-0.90, surpassing DL counterparts (CNN: 0.76-0.79) and aligning with known disease progression patterns. Through rigorous patient-level cross-validation, REN demonstrates strong generalizability and clinical interpretability, presenting a scalable, anatomically-guided approach readily extensible to other structured medical imaging applications.

Figures (4)

• Figure 1: Overview of the REN (Regional Expert Networks) framework. (A) Anatomical region extraction: preprocessing and lobe segmentation assign CT regions to seven experts (five lobes plus bilateral lungs). (B) Individual expert training: CNN, ViT, Mamba, and radiomics (XGBoost) experts are trained on masked inputs with validation AUCs recorded. (C) Gating function extraction: dynamic weighting strategies (performance-, feature-, and learned-based) are applied to expert outputs. (D) End-to-end MoE integration: expert weights and global SwinUNETR features are fused for patient-level ILD classification.
• Figure 2: Anatomical region extraction pipeline showing the process of generating masked inputs for each of the seven lung regions. The original CT scan is processed with lobe-specific segmentation masks to create region-focused inputs that enable expert specialization.
• Figure 3: Mean AUC per lung region across folds for each architecture. Radiomics experts achieved the highest regional performance, particularly in the lower lobes, while CNN followed closely behind. Mamba and ViT showed more variability across regions.
• Figure 4: Radiomics feature category ablation analysis showing (left) performance using individual feature categories across all lung regions and (right) feature importance measured by AUC drop when each category is removed. Texture features demonstrate particular strength in lower lobe regions, with GLCM achieving 0.828 AUC in Left Lower Lobe and GLRLM reaching 0.845 in Right Lower Lobe.