Automated Deep Learning Protocol for Quantifying Cell Viability in Live/Dead Fluorescence Microscopy

The assessment of live/dead cells by means of fluorescence microscopy is a standard technique for assessing cell viability but relies on manual counting, which is laborious and prone to operator bias. Here we present a step-by-step protocol for automated cell viability quantification from two-channel live/dead fluorescence images using a deep learning pipeline built on a U-Net architecture with a ResNet-50 encoder pretrained on ImageNet. The protocol is organized into two tracks: a no-code inference track that applies a pre-trained model to standard live/dead images and an adaptation track for generating custom annotations and retraining the model for new cell types or staining protocols. The pipeline integrates percentile-based normalization, centroid-based (weakly supervised) annotation, patch-based transfer learning, and distance-transform watershed post-processing for instance-level counting. On an independent 25-image test set of human dermal fibroblasts, the model achieved counting R² values of 0.986 (live) and 0.999 (dead) and a mean absolute viability error of 0.61 percentage points, providing a reproducible framework adaptable to diverse cell types and imaging platforms.