MoPET: Parameter-Efficient Mixture-of-Experts for Unified Medical Image Classification @ MICCAI 2026 EMA Workshop

[Preprint (coming soon)] [PyPI] [Pretrained Weights (HF)] [Citation]

Overview

MoPET is a mixture-of-experts method for parameter-efficient fine-tuning (PEFT) of foundation models on 2D medical images. A learned sparse router directs each input through a small, top-k subset of low-rank PEFT experts (LoRA and BOFT) injected into a frozen DINOv3 backbone, sharing adapter capacity across datasets while limiting the gradient conflict that arises when heterogeneous domains are trained jointly. On MedMNIST+, this design lets a single model consolidate multiple classification tasks that would otherwise each need their own fine-tuned network.

Standard PEFT trains one adapted model per dataset (left). MoPET routes each input through a shared feature router to a frozen pretrained backbone plus a pool of specialized PEFT experts, consolidating all datasets into a single multi-domain model (right).

(A) Interleaved batch sampling keeps every training batch diverse across the pooled datasets. (B) Inside a MoPET layer, a router selects the top-k PEFT experts per token; their outputs are combined with the frozen pretrained projection. (C) A dataset router uses the pooled feature and the sample's dataset id to dispatch it to the matching classification head.

Key Contributions

• Across 12 MedMNIST+ datasets, parameter-efficient fine-tuning of a frozen foundation backbone outperforms full end-to-end fine-tuning.
• MoPET, a mixture-of-experts built entirely from PEFT modules, unifies distinct classification tasks in a single model and beats isolated per-domain adapters on a four-dataset pool.
• A cross-domain "booster" dynamic where co-training with auxiliary datasets improves accuracy on data-constrained target datasets.

Installation

From PyPI (the distribution is mopet-moe; it still imports as mopet):

pip install mopet-moe

From source (adds the reproduction/training stack):

git clone https://github.com/sdoerrich97/mopet.git && cd mopet
pip install -e ".[experiments]"

Quick Start

from mopet import create_model

model = create_model(weights="unified").eval()   # downloads the adapter weights from HF; frozen DINOv3 from timm

See examples/ for runnable inference, minimal-API, and training-CLI notebooks.

Model Zoo

Weights live in the MoPET HuggingFace Collection. The Identifier column is exactly the string to pass as create_model(weights=...).

Each release carries only the ~7M trainable parameters (PEFT experts, routers, and per-dataset heads); the frozen DINOv3 backbone is reconstructed from timm at load time.

Project Structure

mopet/            # the installable package: MoPET model, MoE/PEFT experts, create_model factory
experiments/      # reproduction code: configs, data, metrics, reference baselines, entry scripts — not on PyPI
examples/         # runnable inference / usage / training notebooks
for_huggingface/  # staged config + model-card assets for the released HF checkpoints
tests/            # pytest suite
scripts/          # local + cluster (SLURM/Apptainer) job generators

Citation

@inproceedings{doerrich2026mopet,
  title     = {{MoPET}: Parameter-Efficient Mixture-of-Experts for Unified Medical Image Classification},
  author    = {Doerrich, Sebastian and W{\"u}rtinger, Daniel and Di Salvo, Francesco and Rai, Shyam Nandan and Ledig, Christian},
  booktitle = {MICCAI 2026 Workshop on Efficient Medical AI (EMA)},
  year      = {2026},
}

License

The mopet code is released under the MIT License. The frozen DINOv3 backbone weights are downloaded separately via timm/HuggingFace under their own upstream license; only the code in this repository and the trainable-parameter checkpoints in the HuggingFace collection above are covered by the MIT license.

Changelog

v0.1.0

• Initial public release accompanying the MICCAI 2026 EMA Workshop paper.