Fine-Grained Hard-Negative Mining: Generalizing Mitosis Detection with a Fifth of the MIDOG 2022 Dataset

Making histopathology image classifiers robust to a wide range of real-world variability is challenging task. Here, we describe candidate deep learning solution for the Mitosis Domain Generalization Challenge 2022 (MIDOG) address problem generalization mitosis detection in images hematoxylin-eosin-stained histology slides under high (scanner, tissue type and species variability). Our approach consists training rotation-invariant model using aggressive data augmentation with set enriched hard negative examples automatically selected from unlabeled part challenge dataset. To optimize performance our models, investigated mining regime search procedure that lead us train best subset patches representing 19.6% partition ensemble achieved $$\textrm{F}_{1}$$ -score .697 on final test after automated evaluation platform, achieving third overall score MIDOG Challenge.