Dividing cells: Automatic classification of mitotic events

Abstract

Manual annotation of overnight cell imaging datasets is a laborious process prone to user error. Automating the process would provide an efficient and consistent alternative. Training data was generated from imaging datasets through segmentation and annotation. The data was then processed and split into different sets. Classifiers were trained and compared on multiple validation metrics. Best performing models were subsequently validated on novel data. We found that models trained on class-balanced data performed better than those trained on full class data. We also found that applying feature reduction techniques further increases model performance, especially in combination with class balancing. Feature reduction through GCV-based estimate of error consistently increased performance, while cross-correlation based feature reduction only did so in specific instances. Through feature reduction, we found that specifically intensity- and size-related features are relevant to mitosis recognition. Lastly, we found that most false predictions in the best performing models were limited to directly before or after true mitotic events.