Improved Segmentation of Confocal Calcium Videos of Hela Cells Using Deep-Learning-Assisted Watershed Algorithm

Area of Science:

• Biomedical imaging
• Computational biology
• Drug discovery

Background:

• Calcium imaging is crucial for high-throughput drug screening but hindered by manual cell identification.
• Automated cell segmentation is needed to unlock the full potential of calcium imaging in drug discovery.

Purpose of the Study:

• To develop a deep learning-enhanced watershed segmentation method for accurate and efficient cell identification in calcium imaging.
• To overcome the limitations of manual cell segmentation in high-throughput screening.

Main Methods:

• A pipeline using two Convolutional Neural Networks (CNNs), modified U-Net and YOLOv5, was trained for cell segmentation.
• YOLOv5 detected cells, with bounding boxes refined using Intersection over Union (IoU) thresholding.
• Watershed transform, guided by refined cell centers, segmented individual cells, accounting for complex morphology and time-course variations.

Main Results:

• The proposed deep learning-enhanced watershed segmentation method significantly improved segmentation accuracy.
• Outperformed existing methods like U-Net, CellPose, and a 2D method (Med IP).
• Effectively handled complex cell morphologies and temporal dynamics in calcium imaging data.

Conclusions:

• Deep learning-enhanced watershed segmentation offers a robust solution for automated cell identification in calcium imaging.
• This method significantly advances the efficiency and accuracy of high-throughput drug screening.
• Enables better utilization of calcium imaging data for drug discovery and development.