Efficient Fine-Tuning of DINOv3 (EFTD) for Atypical vs Normal Mitotic Figure Classification

This repository contains the code, including the training and inference scripts for our winning solution to MIDOG 2025 — Task 2 (Atypical vs Normal Mitotic Figure Classification).

This work was conducted in collaboration between Mines Paris – PSL and Sanofi Digital R&D.

Task description:

• Dataset: H&E images of mitotic figures
• Goal: classify Atypical Mitotic Figures (AMF) vs Normal Mitotic Figures (NMF)
• Input: 128×128 patches, centered on mitoses at 40× magnification
• More details: MIDOG 2025 Task 2 Description

Our method reached 🏆 1st place on the final hidden test set.

🔬 Method Summary

In this work, we fine-tuned the recently published DINOv3-H+ Vision Transformer, pretrained on natural images, using Low-Rank Adaptation (LoRA), training only ~1.3M parameters in total.

You can find in the code:

• Our DINOv3-H+ architecture: _build_vith16plus_dinov3_hf
• Implementation of the Multi-Macenko Augmentation: AugmentMultiMacenkoNormalizer
• Domain-weighted Focal Loss to handle domain and class imbalance in training data: DomainWeightedFocalLoss

Despite the strong shift between natural images and histopathology, the model transfers effectively and achieves state-of-the-art performance.

📥 Installation

Create a conda environment using Python 3.11:

conda create --name dinov3anmf python=3.11
conda activate dinov3anmf
pip install -r requirements.txt

📦 Download MIDOG 2025 Task 2 Data

bash scripts/download_midog_task2.sh <DATA_DIR>

This script will download MIDOG++, Ami-BR, AtNormBr and OMG-Octo datasets in < 10 minutes. Data mostly presented here: https://midog2025.deepmicroscopy.org/datasets/

🚀 Training

Quick launch

bash scripts/launch_training.sh

You may need to adapt data.root_dir inside the bash script (to point to your data directory)

Manual training

python run_train.py <override parameters via Hydra>

Make sure you have access to the pretrained backbones defined in:

FOUNDATION_HF_CKPT_REGISTRY in src/foundation_models.py

🔎 Inference

You can download the trained model weights from the GitHub Release.

Then open inference.ipynb. This notebook shows:

• how to load the DINOv3 backbone
• how to run predictions on new patches.

🙏 Acknowledgements

We would like to thank the MIDOG 2025 organizers — Marc Aubreville, Jingna Qiu, Sweta Banerjee, Katharina Breininger, and the MICCAI Society

    

📄 Citation

If you use this repository or model in your work, please cite:

@InProceedings{balezo2026efficientfinetuningdinov3,
    author="Balezo, Guillaume
    and Bourgade, Rapha{\"e}l
    and Feki, Hana
    and Monnier, Lily
    and Blons, Matthieu
    and Blondel, Alice
    and Decenci{\`e}re, Etienne
    and Planas, Albert Pla
    and Walter, Thomas",
    editor="Aubreville, Marc
    and Bertram, Christof A.",
    title="Efficient Fine-Tuning of DINOv3 Pretrained on Natural Images for Atypical Mitotic Figure Classification",
    booktitle="Mitotic Figure Detection and Atypia Classification in Whole Slide Images",
    year="2026",
    publisher="Springer Nature Switzerland",
    address="Cham",
    pages="15--25",
    isbn="978-3-032-25180-0",
    doi="10.1007/978-3-032-25180-0_2",
    url="https://doi.org/10.1007/978-3-032-25180-0_2"
}