A Weakly Supervised Learning Method for Cell Detection and Tracking Using Incomplete Initial Annotations.

Area of Science:

• Biomedical Imaging
• Computational Biology
• Machine Learning

Background:

• Cell detection in microscopy is crucial for biomedical research but challenging due to dynamic cell appearance and number.
• Current convolutional neural network (CNN) methods require extensive manual annotations, increasing time and cost.
• Weakly supervised learning offers a potential solution to reduce annotation burden.

Purpose of the Study:

• To develop a novel weakly supervised learning framework for cell detection and tracking in microscopy image sequences.
• To train deep neural networks using incomplete initial labels, reducing the need for comprehensive manual annotation.
• To improve the robustness and accuracy of cell detection and tracking in challenging datasets like Induced Pluripotent Stem (iPS) cells.

Main Methods:

• Proposed a weakly supervised learning framework utilizing incomplete cell markers from fluorescent images for initial training.
• Employed iterative label updating by combining detection and tracking results to enhance model robustness.
• Evaluated the framework on iPS cell datasets and the public FluoN2DH-GOWT1 dataset from the Cell Tracking Challenge (CTC).

Main Results:

• Achieved high detection accuracy (DET) scores of 0.862 and 0.924 on iPS cell datasets.
• Demonstrated significant performance improvement on the FluoN2DH-GOWT1 dataset, with DET increasing from 0.130 to 0.903 (10% labels).
• Showcased that performance improved with increased label quality, surpassing fully supervised methods with 60% labels.

Conclusions:

• The developed weakly supervised framework effectively performs cell detection and tracking with incomplete annotations.
• The iterative label refinement strategy enhances model robustness and accuracy.
• This approach offers a more efficient and cost-effective alternative to fully supervised methods in cell imaging analysis.