OPTIMIZING CONTEXTUAL FEATURE LEARNING FOR MITOSIS DETECTION WITH CONVOLUTIONAL RECURRENT NEURAL NETWORKS

Automatic detection of mitosis in cell videos is essential for research in many fields including stem cell biology and pharmacology. Current state-of-the-art graph-based and deep learning models for mitosis detection rely on candidate sequence extraction that locates the mitotic events at the center of the input frame for optimal contextual feature learning. We propose a method to detect mitosis, by extending convolutional long short-term memory (LSTM) neural networks to remove the candidate sequence extraction step. Our method maintains a high detection accuracy by using the entire video frames as the input, instead of small crops from the original frames and this, acts to preserve the complete contextual features of mitotic events. We evaluated our method on a dataset of stem cell phase-contrast microscopy videos. Under conditions of a temporal tolerance of 1 and 3 frames, our method achieved a detection F1-score of 0.880 and 0.911, which outperformed state-of-the-art benchmark methods by approximately 0.15 in F1-score.