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

1. Problem Statement

Context: Interstitial Lung Disease (ILD) encompasses over 200 diverse pulmonary disorders. Accurate diagnosis via High-Resolution Computed Tomography (HRCT) is critical but challenging due to the heterogeneity of disease patterns across different anatomical regions of the lung.
Limitations of Current Approaches:

• Global Analysis: Traditional Deep Learning (DL) models treat the lung as a homogeneous unit, potentially diluting region-specific pathological signals.
• Generic Mixture-of-Experts (MoE): Standard MoE architectures rely on unconstrained, domain-agnostic routing mechanisms that assume homogeneous expert capabilities. They fail to leverage the inherent anatomical structure and physiological constraints of medical imaging, leading to suboptimal specialization and poor clinical interpretability.
• Lack of Integration: Existing fusion strategies often ignore regional anatomical heterogeneity when combining radiomics (handcrafted features) with deep learning.

2. Methodology: Regional Expert Networks (REN)

The authors propose REN, the first anatomically-informed MoE framework designed specifically for medical imaging. The framework operates in four stages:

A. Anatomical Region Extraction & Expert Definition

• Anatomical Priors: Instead of learning implicit partitions, REN explicitly aligns experts with anatomical regions.
• Seven Specialized Experts: The lung is segmented into seven distinct regions: five individual lobes (Left Upper, Left Lower, Right Upper, Right Middle, Right Lower) and two bilateral combinations (Left Lung, Right Lung).
• Masked Inputs: Each expert receives a CT scan masked to its specific anatomical region, forcing the model to learn region-specific pathological patterns rather than global averages.

B. Multi-Modal Expert Architectures

The framework trains specialized "experts" using four distinct backbones for each of the seven regions:

1. 3D CNN: A custom architecture with progressive blocks (24→48→96 filters).
2. 3D Vision Transformer (ViT): Adapting Dosovitskiy's architecture for volumetric data.
3. State-Space Model (Mamba): Utilizing the Mamba architecture for efficient long-range dependency modeling.
4. Radiomics (XGBoost): Extracts 107 quantitative features (texture, shape, intensity) per region using PyRadiomics.

C. Multi-Modal Gating Mechanisms

REN introduces novel gating strategies to dynamically weight expert contributions based on input data:

• Performance-Based: Weights derived from validation AUCs (specifically Radiomics-guided weighting).
• Feature-Based: Weights derived from intermediate feature representations (magnitude, variance, entropy).
• Learned Gating: Lightweight neural networks (MLP or Attention-based) that predict input-dependent weights.
• Hybrid Integration: The final prediction fuses global contextual features (extracted by a SwinUNETR backbone) with the weighted sum of regional expert features.

D. Training Strategy: Staged vs. End-to-End

• Stage 1-2 (Staged): Experts are trained independently on masked inputs. Gating weights are derived from validation performance (e.g., Radiomics AUC) and fixed.
• Stage 3 (End-to-End): A fully differentiable MoE system where backbone, experts, and gating are jointly optimized.
• Key Finding: The authors found that staged ensembling (fixing weights after independent training) outperforms end-to-end joint optimization. Joint training caused "expert collapse," where shared gradients eroded the strict anatomical specialization required for ILD diagnosis.

3. Key Contributions

1. Anatomically-Constrained MoE: A novel framework that embeds domain-specific anatomical constraints directly into the expert routing mechanism, mirroring radiological practice.
2. Multi-Modal Gating: A strategy that integrates traditional radiomics biomarkers with deep learning features to dynamically balance complementary information sources.
3. Comprehensive Architecture Evaluation: Systematic comparison of CNN, ViT, and Mamba architectures across anatomical regions, revealing that CNN and Radiomics excel in lower lobes (where ILD typically initiates), while Transformers offer complementary sensitivity in diffuse regions.
4. Interpretability: The model provides per-lobe contribution maps, allowing clinicians to see exactly which anatomical regions drove the diagnosis, aligning with clinical workflows.

4. Experimental Results

Dataset: Retrospective cohort of 597 patients (1,898 longitudinal HRCT scans) from the Northwestern Scleroderma Registry.
Evaluation: Strict patient-level cross-validation (5-fold) to prevent data leakage.

Performance Metrics:

• Baseline: SwinUNETR (Single Model) achieved an AUC of 0.7685.
• REN (Radiomics-Guided Ensemble): Achieved an average AUC of 0.8646 ± 0.0467.
  • Improvement: +12.5% over the baseline (p = 0.031).
  • Regional Performance: Lower-lobe experts achieved AUCs of 0.88–0.90, outperforming DL baselines (0.76–0.79) and mirroring known patterns of basal ILD progression.
• Comparison: Radiomics-guided MoE consistently outperformed neural-network-only MoEs and end-to-end gated variants.
• Efficiency: Inference time is ~62ms per patient (16 samples/sec), suitable for real-time clinical decision support.

Ablation Insights:

• Staged vs. End-to-End: Staged ensembling (fixed weights) was superior. End-to-end training degraded performance by blurring anatomical specialization.
• Radiomics Role: Radiomics features were not just fused at the representation level but served as reliability signals to guide expert weighting, proving crucial for capturing subtle fibrotic patterns in lower lobes.
• Regularization: Entropy, load-balancing, and diversity regularizers were essential to prevent expert collapse and ensure equitable expert utilization.

5. Significance and Clinical Impact

• Clinical Relevance: The 12.5% AUC improvement is statistically significant and clinically meaningful, potentially enabling earlier detection of ILD, which is critical for slowing disease progression.
• Interpretability: Unlike "black box" models, REN offers spatial attribution. It explicitly identifies which lung lobes contributed to the diagnosis, aligning with how radiologists evaluate scans.
• Scalability: The framework is backbone-agnostic and demonstrates that integrating handcrafted radiomics with deep learning via anatomical constraints yields superior results compared to pure deep learning approaches.
• Future Directions: The authors plan multi-institutional validation, expansion to multi-class ILD subtyping, and knowledge distillation to compress the ensemble for deployment.

Conclusion: REN establishes a new paradigm for medical imaging AI by moving away from generic, unconstrained MoE designs toward anatomically-grounded, region-specific expert systems. It demonstrates that incorporating domain knowledge (anatomy) and multi-modal data (radiomics) into the routing mechanism is essential for high-performance, interpretable medical diagnosis.