Domain-Expert-Guided Hybrid Mixture-of-Experts for Medical AI: Integrating Data-Driven Learning with Clinical Priors

TL;DR

This work addresses data scarcity in medical imaging by introducing Domain-Knowledge-Guided Hybrid MoE (DKGH-MoE), which unifies data-driven MoE routing with domain-expert guidance based on clinician eye-gaze cues. The framework comprises a Data-Driven MoE, a Domain-Expert MoE, and a fusion gate that balances their contributions, enhancing both performance and interpretability. In INBreast experiments, DKGH-MoE consistently outperforms baselines on ResNet backbones under dense and sparse routing, with notable gains in ACC and AUC, and reveals gaze-aligned expert specialization through visualization. The approach demonstrates that incorporating clinical priors into MoE routing can reduce data requirements and align model reasoning with human clinical expertise, offering practical benefits for data-efficient medical AI and future extensions to other tasks and priors.

Abstract

Mixture-of-Experts (MoE) models increase representational capacity with modest computational cost, but their effectiveness in specialized domains such as medicine is limited by small datasets. In contrast, clinical practice offers rich expert knowledge, such as physician gaze patterns and diagnostic heuristics, that models cannot reliably learn from limited data. Combining data-driven experts, which capture novel patterns, with domain-expert-guided experts, which encode accumulated clinical insights, provides complementary strengths for robust and clinically meaningful learning. To this end, we propose Domain-Knowledge-Guided Hybrid MoE (DKGH-MoE), a plug-and-play and interpretable module that unifies data-driven learning with domain expertise. DKGH-MoE integrates a data-driven MoE to extract novel features from raw imaging data, and a domain-expert-guided MoE incorporates clinical priors, specifically clinician eye-gaze cues, to emphasize regions of high diagnostic relevance. By integrating domain expert insights with data-driven features, DKGH-MoE improves both performance and interpretability.

Figures (3)

• Figure 1: Overall architecture of the network with the plug-and-play DKGH-MoE module incorporated.
• Figure 2: Architecture of DKGH-MoE module.
• Figure 3: Top-1 expert assignments. Each row shows six randomly sampled gaze heatmaps routed to the same expert, illustrating expert specialization driven by clinician attention.