Sequential Saliency Guided Deep Neural Network for Joint Mitosis Identification and Localization in Time-Lapse Phase...

Area of Science:

• Biomedical research
• Cell biology
• Microscopy image analysis

Background:

• Accurate analysis of cell mitotic behavior is crucial for biomedical research and diagnostics.
• Automated mitosis detection systems require improved accuracy for identifying and localizing mitotic events.

Purpose of the Study:

• To propose a novel deep neural network, SSG-DNN, for joint identification and localization of mitotic events in time-lapse phase contrast microscopy images.
• To develop a weakly supervised method for mitosis detection using only sequence-wise labels, eliminating the need for complex preconditioning.

Main Methods:

• The sequential saliency guided deep neural network (SSG-DNN) integrates visual context learning, sequential saliency modeling, and sequence structure modeling.
• The method performs end-to-end learning for visual feature extraction and sequential structure modeling.
• It operates in a one-shot manner, independent of preconditioning methods for mitotic candidate extraction.

Main Results:

• The SSG-DNN method demonstrates superior performance in both mitosis identification and localization tasks.
• Evaluated on the C3H10 dataset and the challenging C2C12-16 dataset, the method showed significant improvements.
• It is the first weakly supervised approach to achieve joint mitosis identification and localization using only sequence-wise labels.

Conclusions:

• The proposed SSG-DNN method offers a powerful and efficient solution for automated mitosis detection in microscopy.
• Its ability to jointly identify and localize mitotic events with weak supervision advances the field of automated cell analysis.
• The method's robustness and effectiveness are validated on diverse and challenging datasets.