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

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer-Aided Diagnosis

Background:

• Mixture-of-Experts (MoE) models offer scalable learning through conditional computation.
• Conventional MoE designs lack anatomical specificity, misaligning with medical imaging's regional heterogeneity.
• Interstitial Lung Disease (ILD) classification requires models that account for varied pathological patterns across lung regions.

Purpose of the Study:

• Introduce Regional Expert Networks (REN), an anatomically-informed MoE framework for medical image classification.
• Enable precise modeling of region-specific pathological variations in the lungs.
• Enhance ILD classification accuracy using multi-modal gating and specialized experts.

Main Methods:

• Developed REN, an MoE framework with seven experts dedicated to distinct lung regions.
• Employed multi-modal gating to integrate radiomics biomarkers and deep learning features (CNN, ViT, Mamba).
• Applied REN to classify ILD in a longitudinal cohort of 597 patients and 1,898 scans.

Main Results:

• The radiomics-guided REN ensemble achieved an average AUC of 0.8646 ± 0.0467 for ILD classification.
• REN demonstrated a +12.5% improvement over the SwinUNETR baseline (AUC 0.7685).
• Specialized lower-lobe experts reached AUCs of 0.88-0.90, outperforming DL baselines and aligning with ILD progression patterns.

Conclusions:

• REN establishes a scalable, anatomically-guided framework for medical image classification.
• The model shows strong generalizability and clinical interpretability in ILD classification.
• REN's approach is potentially extensible to other structured medical imaging tasks.