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

Area of Science:

• Cell Biology
• Microscopy Image Analysis
• Computational Biology

Background:

• Accurate identification of mitosis events and their stages is crucial for understanding cell division.
• Time-lapse phase-contrast microscopy generates complex image data requiring sophisticated analysis techniques.

Purpose of the Study:

• To develop and validate a robust system for automated mitosis event localization and stage localization in microscopy images.
• To improve the precision and recall of mitosis event detection and reduce errors in temporal stage classification.

Main Methods:

• A three-step approach combining Low-Rank Matrix Recovery (LRMR) for salient region detection, a Hierarchical Convolution Neural Network (HCNN) for classification using appearance and motion cues, and a Two-stream Bidirectional Long-Short Term Memory (TS-BLSTM) for temporal stage segmentation.
• Utilizing visual appearance and motion cues within candidate patch sequences for accurate mitosis event classification.
• Employing TS-BLSTM to segment detected mitosis events into four distinct temporal stages.

Main Results:

• The integrated system (LRMR, HCNN, TS-BLSTM) demonstrated superior performance in mitosis event localization, achieving 99.2% precision and 98.0% recall.
• The method achieved an average error of 0.62 frames for mitosis stage localization across five challenging image sequences.
• Outperformed existing state-of-the-art methods in both event and stage localization tasks.

Conclusions:

• The proposed computational framework effectively automates the localization of mitosis events and their temporal stages in phase-contrast microscopy.
• This method offers a significant advancement for quantitative analysis of cell division dynamics from microscopy data.
• The high accuracy and performance metrics suggest broad applicability in biological research requiring precise cell cycle analysis.