Cell mitosis event analysis in phase contrast microscopy images using deep learning

In this paper, we solve the problem of mitosis event localization and its stage localization in time-lapse phase-contrast microscopy images. Our method contains three steps: first, we formulate a Low-Rank Matrix Recovery (LRMR) model to find salient regions from microscopy images and extract candidate patch sequences, which potentially contain mitosis events; second, we classify each candidate patch sequence by our proposed Hierarchical Convolution Neural Network (HCNN) with visual appearance and motion cues; third, for the detected mitosis sequences, we further segment them into four temporal stages by our proposed Two-stream Bidirectional Long-Short Term Memory (TS-BLSTM). In the experiments, we validate our system (LRMR, HCNN, and TS-BLSTM) and evaluate the mitosis event localization and stage localization performance. The proposed method outperforms state-of-the-arts by achieving 99.2% precision and 98.0% recall for mitosis event localization and 0.62 frame error on average for mitosis stage localization in five challenging image sequences.