MoPET's mixture-of-experts boosts medical image accuracy to 93.46%

Training a generalist medical imaging AI is hard: lung X-rays, retinal scans, and pathology slides demand different visual features, and lumping them into one network often triggers negative transfer—where conflicting gradients degrade performance. Parameter-efficient fine-tuning (PEFT) avoids overfitting by freezing a foundation model and inserting small adapters, but the standard approach trains a separate adapter per task, wasting capacity and blocking cross-domain knowledge sharing. Enter MoPET, developed by Sebastian Doerrich and colleagues, which merges these adapters into a mixture-of-experts (MoE) architecture with a learned sparse router. Each input activates only a subset of low-rank PEFT experts, letting the model share capacity across datasets while isolating gradient updates to minimize conflict—essentially getting the best of both specialization and generalization.

On the MedMNIST benchmark, MoPET delivers concrete gains. First, the team confirmed PEFT outperforms full end-to-end fine-tuning, lifting average accuracy from 86.50% to 88.97%. Then, a single MoPET model consolidating four heterogeneous datasets reached 93.46% average accuracy—up from 92.83% for the best isolated PEFT adapter. Critically, MoPET also handles data-scarce clinical targets: co-training with auxiliary datasets raised target accuracy from 81.58% to 83.58%, showing that shared expertise improves rare-disease classification. Accepted to EMA4MICCAI 2026 and released as open source, MoPET offers a practical roadmap for unified medical AI—fewer models, lower compute, and better accuracy on hard cases.