AI-enabled live-dead cell viability classification and motion forecasting

Distinguishing live from dead cells is crucial in a wide variety of research fields, including regenerative medicine, toxicology, pharmacology, and cellular product manufacturing, because it allows researchers to evaluate the efficacy and toxicity of molecules, materials, and therapies and ensures the quality of manufactured cell products. The cost of failure or uncertainty for live-dead analysis can be particularly high for therapeutic compound screening and evaluating medical therapies. Here, we present a novel deep learning framework that integrates a self-attention UNet for segmentation and a transformer network for dynamic tracking of cell movements. Our proposed model achieves state-of-the-art performance, with a high intersection-over-union (IoU) score of 96% and an area-under-curve (AUC) score of 99% for cell segmentation and over 65% IoU of full image cell motion forecasting, highlighting its ability to predict cell dynamics accurately. The self-attention mechanism significantly enhances the model’s ability to differentiate live and dead cells, even in densely packed or morphologically diverse environments. Additionally, the transformer network effectively captures temporal dependencies, enabling precise predictions of future cell movements. This integrated framework demonstrates robust performance across diverse datasets, consistently outperforming existing methods. By offering high-accuracy segmentation and predictive modeling, our approach provides a transformative tool for advancing cellular analysis in research and clinical applications, including cell therapeutics, cancer diagnostics, drug development, and regenerative medicine.