MIDOG is a challenge organized on the occasion of the MICCAI conference, dedicated to medical image computing and computer-assisted intervention. This bioinformatics competition focused on mitotic figures, which are a key biomarker in tumor grading. The variability of clinical samples makes the detection and classification of mitotic figures difficult for current AI models. The aim of MIDOG 2025 was to address this problem by testing algorithms in atypical conditions.

With their solutions, Thomas Walter’s team won:

• 2nd place in Task 1 focused on mitotic figures detection. They designed a robust and efficient YOLOv12 one-stage model, combined with a streamlined preprocessing pipeline. The approach enables fast detection of both mitotic figures and hard negatives, and benefits from a precise data augmentation process integrating multi-target Macenko stain normalization.
• 1st place in Task 2 dedicated to atypical mitotic figure classification. They leveraged the DINOV3-H+ foundation model, originally pretrained on natural images; and fine-tuned it with LoRA, requiring only ~1.3M trainable parameters. To tailor the model to histopathology, they incorporated extensive domain-specific data augmentations and implemented a domain-aware focal loss, allowing them to better handle both class and domain imbalance during training.

You can read the preprint of their solutions here:

• Task 1: https://www.arxiv.org/abs/2509.02593 (update soon)
• Task 2: https://arxiv.org/abs/2508.21041