EFTD-MIDOG-AMF

Counting mitoses is routine in tumour grading, but the subtype matters as much as the count. Atypical mitotic figures — cell divisions with abnormal spindle geometry or chromosome distribution — signal genomic instability and correlate with poor outcome, yet they are scarce, morphologically subtle, and scored inconsistently even by trained pathologists. The MIDOG 2025 challenge added a dedicated benchmark for this problem: given a 128x128 pixel crop centred on a mitotic figure, classify it as atypical or normal, across tumour types, species, scanners, and laboratories the model has never seen.

EFTD-MIDOG-AMF is the winning entry for that task, from Mines Paris – PSL and Sanofi working with Institut Curie. Rather than building on a histopathology foundation model, the authors adapt DINOv3-H+ — a vision transformer self-supervised on the LVD-1689M natural-image corpus — using low-rank adaptation, training roughly 1.3 million parameters while the backbone stays frozen. Paired with aggressive stain augmentation and a loss reweighted by training domain, the adapted model took first place on the challenge's held-out final test set.

The result is a concrete data point in an open argument: whether specialised pathology pretraining is necessary, or whether a strong general-purpose visual backbone plus a few million task parameters is enough. Here the generic backbone won, and it did so on the challenge explicitly designed to punish domain overfitting.

#Key Features

• Low-rank adaptation of a frozen backbone: LoRA with rank 8 and alpha 16 is applied only to the query and value projections of the attention layers, with 0.05 dropout, leaving about 1.3M trainable parameters plus a linear head on the class token.
• Domain-weighted focal loss: focal loss (alpha 0.25, gamma 2) is combined with per-domain weights set to the inverse square root of domain size and capped at a 3:1 ratio, addressing both the roughly 20% atypical class imbalance and heterogeneous domain sizes.
• Multi-Macenko stain augmentation: stain matrices extracted from ten images per domain — drawn from the challenge data plus MITOS CMC, MITOS CCMCT, and TCGA COAD/BLCA — are averaged into randomised normalisation targets, then perturbed in stain space to widen the staining variation seen during training.
• Rotation test-time augmentation: logits are averaged over four rotated views at inference.
• Released adapter and head: the trained LoRA weights and classification head ship as a GitHub release alongside training scripts and an inference notebook, so the classifier can be run without reproducing the training pipeline.

#Technical Details

Training combined 16,398 mitotic figure crops (12,724 normal, 3,674 atypical) from four sources: the MIDOG 2025 atypical training set (11,939), the TUPAC16 portion of AMi-Br (1,999), AtNorM-Br (711), and OMG-Octo (1,752), all at 128x128 pixels and 0.25 micrometres per pixel. Optimisation used AdamW at learning rate 1e-4 with weight decay 0.1, a cosine schedule with linear warmup over the first 10% of steps, gradient clipping at 1.0, batch size 16, and mixed precision, for 60 epochs on the full combined set.

Balanced accuracy was 0.9485 (± 0.0038) under four-fold cross-validation, 0.8475 on the held-out AMi-Br TUPAC16 external set, 0.9045 on the preliminary test set, and 0.9079 on the final test set of 120 cases spanning 12 human and veterinary tumour types — first place. A fully fine-tuned DINOv3-ConvNeXt-Tiny baseline with roughly 21 times more trainable parameters scored 0.9352 in cross-validation and 0.8085 on AMi-Br, overfitting the training domain more sharply. Swapping plain focal loss for the domain-weighted variant raised cross-validation balanced accuracy from 0.9284 to 0.9341 and the preliminary test score from 0.8870 to 0.9045, but left AMi-Br essentially unchanged, since reweighting can only balance domains that appear in training.

#Applications

The model slots in downstream of mitosis detection: a detector proposes candidate figures across a whole-slide image, and this classifier subtypes each one, producing atypical-mitosis counts and ratios for prognostic assessment. It targets research and challenge use rather than clinical deployment, and the released code registers several histopathology foundation-model backbones — including UNI and Virchow — as drop-in alternatives, making the repository usable as a comparison harness for mitosis subtyping.

#Impact

The entry's significance is the demonstration that a natural-image self-supervised backbone, adapted with a few million parameters, can beat purpose-built approaches on a multi-domain histopathology benchmark. Three limitations bound that claim. The model covers one binary task and has not been validated beyond the MIDOG 2025 benchmark. Code and weights are released under a Sanofi non-commercial licence restricted to academic research and non-profit use, so commercial reuse requires separate permission. And DINOv3-H+ is expensive at inference, an awkward fit for a pipeline that must score hundreds of thousands of candidate patches per slide; the authors point to distillation as the obvious next step.