Balanced Soft mixture-of-expert model for Glaucoma Detection

Glaucoma detection often relies on multiple imaging modalities, such as fundus photographs and optical coherence tomography (OCT). However, training models on these diverse inputs often leads to 'modality imbalance,' where one modality dominates the learning process, causing other informative data sources to be under-optimized. The authors investigate how to create a more robust multi-modal framework that ensures equitable contributions from all input types to improve diagnostic accuracy.

The researchers introduce a Balanced Soft Mixture-of-Experts (SMoE) model. This architecture consists of three specialized experts: two uni-modal experts (each processing a single imaging type) and one multi-modal expert (processing both). A CNN-based gating network dynamically assigns weights to these experts for every input. To ensure that no single expert dominates the learning, the authors incorporate a coefficient-of-variation-based load balancing loss, which forces the model to utilize all experts effectively during training.

The SMoE model was evaluated against several baselines, including standard multi-modal models and existing gradient-modulation techniques (like OPM and OGM), across three datasets: FairVision, FairDomain, and HarvardGF. The proposed SMoE consistently achieved the highest Area Under the Curve (AUC) scores across all three datasets. The results demonstrate that the soft routing mechanism and load-balancing loss effectively mitigate the issues of modality dominance, leading to more stable and accurate predictions compared to conventional approaches that treat all modalities with a uniform learning objective.

By addressing the inherent imbalance in multi-modal medical imaging, this research provides a more reliable framework for automated glaucoma screening. The model's ability to adaptively weight different imaging sources makes it more robust to the noise and variability common in clinical settings, potentially leading to earlier and more accurate diagnosis of irreversible vision loss. Furthermore, the authors suggest the architecture is generalizable to other retinal diseases, such as diabetic retinopathy.