MoPET: Parameter-Efficient Mixture-of-Experts for Unified Medical Image Classification

Researchers introduced MoPET, a parameter-efficient mixture-of-experts method for medical image classification.

It routes inputs through a small set of low-rank experts inside a frozen foundation model, reducing the need to retrain the full network.

On MedMNIST, MoPET outperformed full fine-tuning and standalone adapters, while also consolidating multiple datasets into a single stronger model.

The study also found that auxiliary datasets can improve performance on data-scarce clinical tasks.