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

The potential of calcium imaging in high-throughput drug screening experiments remains underutilized, primarily because of time-intensive manual identification of cells.To overcome these challenges, we propose to use deep learning enhanced watershed segmentation, wherein the complex morphology and the distinctions in the time courses are appropriately accounted through enhanced watershed and multi-frame processing respectively. Our segmentation pipeline involved training of two CNNs (modified U-Net and YOLOv5), one to predict distance transform and other to detect individual cells on the predicted distance transform images. The bounding boxes detected by the YOLO were continuously refined every n frames using an Intersection over Union (IoU) threshold. The centers of the final updated bounding boxes were then used to guide the watershed transform, which segmented the image into individual cells. The proposed method outperformed U-Net (pixel wise classification), CellPose (generalistic deep learning-based software) and 2D method (Med IP) with a significant improvement in segmentation accuracy.