Deeptosis: A Deep Learning-Based Platform for Label-Free Discrimination of Apoptosis and Pyroptosis from Brightfield Microscopy.

MOTIVATION: Accurately distinguishing apoptosis from pyroptosis is essential for studying regulated cell death and its roles in immunity and disease, but their similar morphologies and shared upstream signals make label-free bright-field discrimination difficult. RESULTS: We present Deeptosis, an end-to-end deep learning pipeline that performs automatic cell segmentation (Cellpose) and single-cell classification with a Vision Transformer (ViT). Trained on 26,565 manually annotated bright-field cells (apoptosis, pyroptosis, other), the model achieved a mean AUROC of 0.999 in five-fold cross-validation and retained high performance on an independent test set (AUROC 0.990 apoptosis, 0.982 pyroptosis, 0.983 other). The system outputs color-coded visualization and a per-cell CSV containing coordinates, labels, and confidence scores, and can be operated through a web interface for batch analysis. AVAILABILITY: Source code and scripts are available at GitHub (https://github.com/Bamba-WangLab/Deeptosis); a prototype web app (http://modinfor.com/Deeptosis) demonstrates the workflow. Conclusion Deeptosis provides a label-free framework for quantitative analysis of cell death modalities.