Train for Stain: Adapting for Diverse H&E Staining Profiles Across Centers in Classification of Mitotic Figures (Glioma-MDC 2025)

Mitotic figure detection is an important aspect of grading for many forms of cancer however, there can be very high interrater variability. With the emergence of computational digital pathology, there is great potential for this classification task to be carried out objectively and consistently by a data-driven approach. However, this process is somewhat hampered by the diversity of cellular characteristics across cancer sites and the size of the domain shift between datasets originating from different centers. Building upon a previously developed classifier for H&E stained samples (HPVNet), this study aims to see if the established practices to minimize the impact of domain shift by stain variation can be minimized through a range of augmentation approaches and fine-tuning. The developed model obtained a very high ROC AUC measure (0.99 in train and validation) as well as good accuracy (0.94 in validation) and F1 score (0.94 in validation). The findings show potential but more work is needed to establish a truly generalizable approach.