Attention based Transformers for Mitosis Domain Generalization in H&E stained WSIs

Digital pathology is increasingly being used in the diagnosis and prognosis of tumors. Detection of increased mitosis is a key component of tumor prognostication for various tumors. While manual interpretation is still the most widely used, it suffers from high variability between pathologists, severely affecting accuracy and reproducibility. Furthermore, this requires significant time and effort from expert pathologists thereby limiting its scalability. Many of the current state-of-the-art methods in digital histopathology are specialized for specific tumors. Such models have limited generalization capability in the case of unseen tissue types and in the case of instrument and staining variations. Hence there is a need for accurate methods for detection of mitotic figures which can generalize well to unseen tissue types and is robust against instrument variability.